Agentic Data Platform 🍓

{
  "name": "Agentic Data Platform \ud83c\udf53",
  "description": "senga, a data platform whose customer is an agent, sells predefined pipelines over your own sources: meeting memory (Google Drive), company map from the wiki (Confluence), retro magazine archive (HTTP files), and over public data weather threshold flags from Open-Meteo, each built, tested on sample data and delivered as a Docker image with MCP tools you run on your own infrastructure with your own values. platform.pipelines lists each one's steps and MCP tools. Any other task over your own documents in Google Drive, Confluence or behind an HTTP file listing is quotable too: senga designs a fact schema for it and assembles a documents pipeline of the same known steps (platform.documents). For anything else over public sources or APIs you hold a key for, describe it in plain language and senga designs, writes, builds and tests a custom MCP server for it. Plans that need none of your secrets can also be hosted as a preview. Quoting is free, payment is x402, and nothing is charged until the image passes its test.",
  "url": "https://senga.ngrok.app/a2a",
  "protocolVersion": "0.3.0",
  "preferredTransport": "JSONRPC",
  "supportedInterfaces": [
    {
      "url": "https://senga.ngrok.app/a2a",
      "protocolBinding": "JSONRPC",
      "protocolVersion": "1.0"
    }
  ],
  "version": "1.0.0",
  "documentationUrl": "https://senga.ngrok.app/openapi.json",
  "provider": {
    "organization": "Keboola",
    "url": "https://www.keboola.com"
  },
  "capabilities": {
    "streaming": false,
    "pushNotifications": false,
    "stateTransitionHistory": false
  },
  "defaultInputModes": [
    "text/plain",
    "application/json"
  ],
  "defaultOutputModes": [
    "application/json",
    "text/plain"
  ],
  "skills": [
    {
      "id": "quote-data-pipeline",
      "name": "Quote a data pipeline over your sources",
      "description": "Describe the data you need in plain language. Requests over your own documents in Google Drive, Confluence or behind an HTTP file listing, and weather threshold flags from Open-Meteo, get a predefined pipeline's fixed plan (platform.pipelines lists their steps and MCP tools); other tasks over the same documents get a generated fact schema on the same steps; anything else over public sources or APIs you hold a key for gets an MCP server designed for it (sources, typed tools, storage, schedule). Either way you get a priced, executable plan with expiry and delivery time. Free. Over A2A (JSON-RPC at url, SendMessage or message/send): send what you need as a text part, or a data part {intent, hints}. The task id is the quote_id; reply on it with a data part {answers} while it is input-required. A completed task carries the priced quote as its artifact; buy it over HTTP + x402 as in how_to_buy.",
      "tags": [
        "data pipeline",
        "documents",
        "mcp server",
        "code generation",
        "etl",
        "x402",
        "quote"
      ],
      "examples": [
        "Track the ECB euro reference rates every day, keep the history, and give me MCP tools for the latest rate of a currency and its change over a week.",
        "Earthquakes above magnitude 4 worldwide from USGS, refreshed hourly, with an MCP tool to list the ones near a city.",
        "{\"intent\": \"hourly temperature for Oslo, flag readings below 5 C\", \"hints\": {\"storage\": \"file\"}}"
      ],
      "inputModes": [
        "text/plain",
        "application/json"
      ],
      "outputModes": [
        "application/json"
      ]
    },
    {
      "id": "buy-tested-image",
      "name": "Buy the quoted pipeline as a tested Docker image",
      "description": "Pay a quote with x402 (USDC on eip155:8453 or eip155:84532) at POST https://senga.ngrok.app/v1/orders; the platform builds the image, runs it on sample data and settles only after the test passes. Delivered as an OCI image (registry pull or image.tar) with a manifest and test report, plus a hosted MCP endpoint when asked.",
      "tags": [
        "x402",
        "docker",
        "oci",
        "mcp server"
      ],
      "examples": [
        "{\"quote_id\": \"q_\u2026\", \"delivery\": \"hosted\"}"
      ],
      "inputModes": [
        "application/json"
      ],
      "outputModes": [
        "application/json"
      ]
    }
  ],
  "platform_url": "https://senga.ngrok.app",
  "how_to_buy": [
    "1. POST https://senga.ngrok.app/v1/quotes {\"intent\": \"...\"} (free). If status is needs_input, POST again with {\"quote_id\", \"answers\": {<question id>: <value>}}. A quote never asks about usage: run_cost_estimate (first run, later runs, total) takes the data amount per run (data_items), frequency (interval_seconds) and run count (run_count) from your intent, or from optional numeric answers, and otherwise assumes the source's typical data and 30 days of runs at the plan's frequency (run_cost_estimate.assumed names them). A priced quote is final, but POST {\"quote_id\", \"answers\"} with only data_items, interval_seconds or run_count recalculates run_cost_estimate; price, plan and expiry stay. A document pipeline (Google Drive folder, Confluence site and space, HTTP archive listing) is quoted without waiting for its source: the image reads the source from env at docker run. quote.source_setup shows which source values the intent, hints or answers stated (DRIVE_FOLDER_ID, CONFLUENCE_BASE_URL, CONFLUENCE_SPACE_KEY, SOURCE_URL; they fill sample_config) and lists the open ones as optional_questions: answer (or correct) them with {\"quote_id\", \"answers\"} on the priced quote if you know them, or set them at docker run. Credentials (source_setup.at_docker_run) are never asked.",
    "2. POST https://senga.ngrok.app/v1/orders {\"quote_id\"} \u2192 402 with PAYMENT-REQUIRED; resend the same body with PAYMENT-SIGNATURE (x402 v2, exact, USDC).",
    "3. A valid authorization answers 202 with order_id, access_token, eta_seconds and status_url. No money has moved yet; a replayed payment returns the same order, never a second charge. Every URL in the answer carries ?access_token=, so a client that cannot set headers just fetches it; X-Access-Token or Authorization: Bearer work too.",
    "4. Poll GET status_url every few seconds until status is delivered, failed, payment_required or topup_required (or open GET /v1/orders/{order_id}/events?access_token=\u2026 as Server-Sent Events). Phases: accepted \u2192 building \u2192 testing \u2192 ready \u2192 settling \u2192 delivered. Settlement happens only after the image passes its test; failed means nothing was charged. A 502 or 503 while polling (possibly the proxy's HTML page while senga restarts or deploys, platform_unavailable) is transient: polling is idempotent, so repeat the same GET after a few seconds.",
    "5. When delivered: GET artifact_url (manifest, test report, digest, and ready pull and load commands; the commands read the token from $SENGA_ACCESS_TOKEN, so export SENGA_ACCESS_TOKEN=<access_token> first and keep it out of them). Get the image preferably from the registry: echo \"$SENGA_ACCESS_TOKEN\" | docker login senga.ngrok.app -u <order_id> --password-stdin, then docker pull manifest.image.ref (manifest.image.registry.pinned pins the digest); or, as a fallback, download image.tar (its sha256 is manifest.image.archive_sha256). runner_image is the image block of an evaluation RunSpec, ready to copy: prefer from_registry {ref: the pinned digest, registry_auth: inline login with the access_token}, which pulls layer by layer and resumes an interrupted download; from_archive {archive: {url, sha256 of image.tar, headers: {Authorization: Bearer <access_token>}}} is the fallback, one file that restarts from the beginning when the connection drops; runner_image is secret: it carries the access_token, so hand it to the runner without logging it. manifest.image.digest is the image ID, not the archive hash; manifest.required_env names the settings the image refuses to start without, manifest.domain_tools the tools besides describe, run_now and run_status; manifest.run_with is the image reference a runner starts itself (Agnes adp_run_image): docker pull it when manifest.image.registry is set, otherwise docker load image.tar first, then docker run it with manifest.env on manifest.ports.http; it equals manifest.image.ref. Then export the manifest.env settings without a default in your shell and run manifest.run (it loads image.tar when present, passes those settings with -e and starts manifest.image.ref on the order's own volume adp-data-<order_id>; the image refuses storage that holds another plan's data) on your infrastructure with your own values. The image needs nothing from this platform after that.",
    "6. If status is payment_required (a custom build can outlast a short authorization), the status answer carries pay_again: POST /v1/orders again with the same quote_id (or POST /v1/orders/{order_id}/payment, which takes the order's quote and answers 409 not_waiting_for_payment once the order is past payment_required) and a fresh PAYMENT-SIGNATURE for one entry of pay_again.accepts before pay_again.until (the quote's expiry); the tested image is settled and delivered without rebuilding. After that the order fails with payment_window_closed and nothing is charged. Each quote is ordered at most once.",
    "Optional: add \"delivery\": \"hosted\" to the order body to also get a hosted MCP endpoint at mcp.url (JSON-RPC POST) as a preview. It is offered only for plans that need none of your secrets (see the quote's order.hosting); otherwise the order answers 400 hosting_needs_buyer_secrets before any payment. It runs for 6 hours after delivery (mcp.expires_at); then the container stops, hosting.status becomes expired and mcp.url answers 410 hosting_expired, while the image stays available to run yourself. When every hosted slot is in use the order answers 503 hosting_full before any payment: omit delivery or retry after retry_after_seconds. Without delivery the order is self_hosted: the default when delivery is omitted: we start nothing; get the tested image (docker pull manifest.image.ref after docker login with the order when manifest.image.registry is set, otherwise image.tar from artifact_url) and run manifest.run on your infrastructure with your own values.",
    "7. Lost the access_token? POST /v1/orders/{order_id}/recover {} \u2192 a one-time challenge (5 min). Sign its message with the wallet that paid (EIP-191 personal_sign) and POST {challenge_id, signature} to the same URL \u2192 a new access_token; the old one stops working. POST /v1/orders/recover with {order_id} or {order_id, challenge_id, signature} does the same. Free; no other wallet can recover the order.",
    "8. MCP-only agents buy the same way over MCP at https://senga.ngrok.app/mcp (streamable HTTP): tools describe_platform, quote, get_quote, place_order (x402 over MCP: payment in _meta[\"x402/payment\"], the PaymentPayload object or its base64 PAYMENT-SIGNATURE string), order_status, get_artifact, pay_topup (a topup_required order, paid like place_order), cancel_order (a topup_required order, free) and recover_access; describe_platform explains the steps.",
    "9. If status is topup_required, the work needs more than quoted (e.g. a generated server needs more attempts than the price covers). The status answer carries topup: extra_usd, total_usd, new_delivery_seconds, reason and expires_at. Pay: POST topup.pay.url with a fresh PAYMENT-SIGNATURE for one entry of topup.pay.accepts at the new total; it replaces the original authorization, nothing is settled until the image passes its test. Or cancel: POST topup.cancel.url, which costs nothing. Unanswered offers expire as topup_declined, also free. To see this path on purpose, order scenarios.topup_required: its first attempt fails and the order pauses here.",
    "10. To check that a failed build costs nothing, order the published failing build: scenarios.failing_build gives the quote body and what the order shows (failed, test_failed, no settlement)."
  ],
  "platform": {
    "sources": [
      {
        "type": "google-drive",
        "pipeline": "meetings",
        "credentials": [
          {
            "env": "DRIVE_FOLDER_ID",
            "required": true,
            "secret": false,
            "description": "Folder to read: transcripts (Google Docs, .txt, .vtt) and audio-only recordings directly in it.",
            "example": "<ID of the Google Drive folder with the meeting transcripts>"
          },
          {
            "env": "GOOGLE_SA_JSON",
            "required": true,
            "secret": true,
            "description": "Google Cloud service account key (JSON) the folder is shared with as Viewer; used only for the Drive API (drive.readonly).",
            "example": "<service account JSON with read access to the folder>"
          }
        ],
        "credentials_schema": {
          "type": "object",
          "properties": {
            "DRIVE_FOLDER_ID": {
              "type": "string",
              "description": "Folder to read: transcripts (Google Docs, .txt, .vtt) and audio-only recordings directly in it.",
              "examples": [
                "<ID of the Google Drive folder with the meeting transcripts>"
              ]
            },
            "GOOGLE_SA_JSON": {
              "type": "string",
              "description": "Google Cloud service account key (JSON) the folder is shared with as Viewer; used only for the Drive API (drive.readonly).",
              "examples": [
                "<service account JSON with read access to the folder>"
              ],
              "writeOnly": true
            }
          },
          "required": [
            "DRIVE_FOLDER_ID",
            "GOOGLE_SA_JSON"
          ]
        },
        "runtime_ready": true
      },
      {
        "type": "confluence",
        "pipeline": "wiki",
        "credentials": [
          {
            "env": "CONFLUENCE_BASE_URL",
            "required": true,
            "secret": false,
            "description": "Base URL of the Confluence Cloud site.",
            "example": "<base URL of the Confluence Cloud site, e.g. https://your-site.atlassian.net/wiki>"
          },
          {
            "env": "CONFLUENCE_SPACE_KEY",
            "required": true,
            "secret": false,
            "description": "Space whose pages and attachments are read.",
            "example": "<key of the Confluence space, e.g. ENG>"
          },
          {
            "env": "CONFLUENCE_EMAIL",
            "required": true,
            "secret": false,
            "description": "Atlassian account used with the API token (basic auth).",
            "example": "<e-mail of the Atlassian account that owns the API token>"
          },
          {
            "env": "CONFLUENCE_API_TOKEN",
            "required": true,
            "secret": true,
            "description": "API token of that account; read access to the space's pages and attachments is enough.",
            "example": "<Atlassian API token with read access to the space>"
          }
        ],
        "credentials_schema": {
          "type": "object",
          "properties": {
            "CONFLUENCE_BASE_URL": {
              "type": "string",
              "description": "Base URL of the Confluence Cloud site.",
              "examples": [
                "<base URL of the Confluence Cloud site, e.g. https://your-site.atlassian.net/wiki>"
              ]
            },
            "CONFLUENCE_SPACE_KEY": {
              "type": "string",
              "description": "Space whose pages and attachments are read.",
              "examples": [
                "<key of the Confluence space, e.g. ENG>"
              ]
            },
            "CONFLUENCE_EMAIL": {
              "type": "string",
              "description": "Atlassian account used with the API token (basic auth).",
              "examples": [
                "<e-mail of the Atlassian account that owns the API token>"
              ]
            },
            "CONFLUENCE_API_TOKEN": {
              "type": "string",
              "description": "API token of that account; read access to the space's pages and attachments is enough.",
              "examples": [
                "<Atlassian API token with read access to the space>"
              ],
              "writeOnly": true
            }
          },
          "required": [
            "CONFLUENCE_BASE_URL",
            "CONFLUENCE_SPACE_KEY",
            "CONFLUENCE_EMAIL",
            "CONFLUENCE_API_TOKEN"
          ]
        },
        "runtime_ready": true
      },
      {
        "type": "http-files",
        "pipeline": "http_archive",
        "credentials": [
          {
            "env": "SOURCE_URL",
            "required": true,
            "secret": false,
            "description": "HTTP(S) file listing of the archive ({ items: [{ id, revision, sha256, mime_type, download_url }] }); downloads honour ETag / If-None-Match.",
            "example": "<http(s):// URL of the archive's file listing>"
          }
        ],
        "credentials_schema": {
          "type": "object",
          "properties": {
            "SOURCE_URL": {
              "type": "string",
              "description": "HTTP(S) file listing of the archive ({ items: [{ id, revision, sha256, mime_type, download_url }] }); downloads honour ETag / If-None-Match.",
              "examples": [
                "<http(s):// URL of the archive's file listing>"
              ]
            }
          },
          "required": [
            "SOURCE_URL"
          ]
        },
        "runtime_ready": true
      },
      {
        "type": "weather",
        "pipeline": "weather",
        "credentials": [],
        "credentials_schema": {
          "type": "object",
          "properties": {},
          "required": []
        },
        "runtime_ready": true
      },
      {
        "type": "https_api",
        "provider": "any https API, feed or page, designed per order",
        "description": "https sources fixed at design time: public ones, and APIs that take a static API key or token in a header or query parameter, optionally on your own origin (e.g. your tenant); you inject the values at docker run, each credential is sent only to its source's host and never reaches senga. OAuth, interactive logins, cookies and request signing are not supported",
        "credentials": "declared per order in the quote's config_schema",
        "credentials_schema": {
          "type": "object",
          "description": "Per order: UPPER_SNAKE_CASE env variables from the quote's config_schema; secret ones (writeOnly) are API keys or tokens sent only to the host named in the quote's credentials, others are your https origin for a source",
          "additionalProperties": {
            "type": "string"
          }
        },
        "runtime_ready": true
      }
    ],
    "pipelines": [
      {
        "route": "meetings",
        "title": "UC2 meeting memory (Google Drive)",
        "source_type": "google-drive",
        "steps": [
          "files.sync",
          "document.normalize",
          "facts.extract",
          "facts.validate"
        ],
        "tools": [
          "search_meetings",
          "decisions_about",
          "actions_by_owner",
          "actions_about",
          "actions_without_due_date",
          "revision_changes",
          "meeting_facts"
        ],
        "storage": [
          "file",
          "postgres"
        ],
        "also_injected": [
          {
            "env": "ANTHROPIC_API_KEY",
            "required": false,
            "secret": true,
            "description": "Buyer's Anthropic API key; when set, the image calls claude-sonnet-5-5 at effort low directly to extract decisions and action items, otherwise through senga (optional fallback). Only plans that extract facts get this fallback. With the buyer's own ANTHROPIC_API_KEY the image never calls senga; the order key stays in plan.json unused. Without it, the image calls senga's POST /v1/extract with that key. The key is a metered extraction credential for this order only, not a buyer secret and not a key to the model provider: it can spend only this order's extraction budget (the quote's extraction.budget_usd), and senga stops accepting calls once that budget is spent. Facts and normalized content stay in the buyer's storage; a call sends senga only the documents' text segments with their titles, dates and speakers. run_status shows extract_path and llm_usage.",
            "example": "<buyer's Anthropic API key>"
          }
        ],
        "runtime_ready": true,
        "runtime_missing": {
          "source_types": [],
          "step_kinds": [],
          "tools": []
        },
        "status": "quoted as a plan of known steps",
        "promised_seconds": 900,
        "first_run_usd": "1.81",
        "per_run_cost_estimate_usd": "0.19",
        "cost_model": {
          "itemUnit": "meeting transcript",
          "itemUnits": "meeting transcripts",
          "typicalItems": 60,
          "usdPerRun": "0.01",
          "usdPerItem": "0.030",
          "laterRunShare": 0.1,
          "basis": "~5k tokens in and ~1k out per transcript, audio transcribed in the image (whisper.cpp), on a Sonnet-class model at $3 / $15 per million input / output tokens, billed to the buyer's key; later runs only process new or changed revisions"
        }
      },
      {
        "route": "wiki",
        "title": "UC3 company map from the wiki (Confluence)",
        "source_type": "confluence",
        "steps": [
          "files.sync",
          "document.normalize",
          "facts.extract",
          "facts.validate",
          "facts.reconcile"
        ],
        "tools": [
          "list_sources",
          "get_source",
          "quarantine",
          "entities",
          "describe_entity",
          "who_owns",
          "relations",
          "conflicts",
          "stale_facts"
        ],
        "storage": [
          "postgres",
          "file"
        ],
        "also_injected": [
          {
            "env": "ANTHROPIC_API_KEY",
            "required": false,
            "secret": true,
            "description": "Buyer's Anthropic API key; when set, the image calls claude-sonnet-5-5 at effort low directly to extract entities and relations, otherwise through senga (optional fallback). Only plans that extract facts get this fallback. With the buyer's own ANTHROPIC_API_KEY the image never calls senga; the order key stays in plan.json unused. Without it, the image calls senga's POST /v1/extract with that key. The key is a metered extraction credential for this order only, not a buyer secret and not a key to the model provider: it can spend only this order's extraction budget (the quote's extraction.budget_usd), and senga stops accepting calls once that budget is spent. Facts and normalized content stay in the buyer's storage; a call sends senga only the documents' text segments with their titles, dates and speakers. run_status shows extract_path and llm_usage.",
            "example": "<buyer's Anthropic API key>"
          }
        ],
        "runtime_ready": true,
        "runtime_missing": {
          "source_types": [],
          "step_kinds": [],
          "tools": []
        },
        "status": "quoted as a plan of known steps",
        "promised_seconds": 1200,
        "first_run_usd": "7.37",
        "per_run_cost_estimate_usd": "0.75",
        "cost_model": {
          "itemUnit": "page or attachment",
          "itemUnits": "pages or attachments",
          "typicalItems": 230,
          "usdPerRun": "0.01",
          "usdPerItem": "0.032",
          "laterRunShare": 0.1,
          "basis": "~3k tokens in and ~1.5k out per page or attachment, on a Sonnet-class model at $3 / $15 per million input / output tokens, billed to the buyer's key; later runs only process new or changed revisions"
        }
      },
      {
        "route": "http_archive",
        "title": "UC4 retro magazine archive (HTTP files)",
        "source_type": "http-files",
        "steps": [
          "files.sync",
          "document.normalize",
          "basic.listing",
          "basic.validate",
          "records.dedupe"
        ],
        "tools": [
          "programs",
          "listing",
          "failed_listings",
          "articles",
          "revisions"
        ],
        "storage": [
          "file",
          "postgres"
        ],
        "also_injected": [],
        "runtime_ready": true,
        "runtime_missing": {
          "source_types": [],
          "step_kinds": [],
          "tools": []
        },
        "status": "quoted as a plan of known steps",
        "promised_seconds": 900,
        "first_run_usd": "0.03",
        "per_run_cost_estimate_usd": "0.01",
        "cost_model": {
          "itemUnit": "archive item",
          "itemUnits": "archive items",
          "typicalItems": 120,
          "usdPerRun": "0.01",
          "usdPerItem": "0.0002",
          "laterRunShare": 0.1,
          "basis": "everything is deterministic, no LLM tokens: masthead, classification, genre, listing extraction, target by dialect and validation take up to ~10 s of CPU per item for the PDF text layer or tesseract OCR with a stronger pass on unreadable glyphs, at ~$0.05 per vCPU-hour rounded up for memory and the download; later runs only process new or changed revisions"
        }
      },
      {
        "route": "weather",
        "title": "weather threshold flags from Open-Meteo",
        "source_type": "weather",
        "steps": [
          "weather.fetch",
          "threshold.flag"
        ],
        "tools": [
          "latest_reading",
          "flagged_readings"
        ],
        "storage": [
          "file",
          "postgres"
        ],
        "also_injected": [],
        "runtime_ready": true,
        "runtime_missing": {
          "source_types": [],
          "step_kinds": [],
          "tools": []
        },
        "status": "quoted as a plan of known steps",
        "promised_seconds": 90,
        "first_run_usd": "0.00",
        "per_run_cost_estimate_usd": "0.00",
        "cost_model": {
          "itemUnit": "city",
          "itemUnits": "cities",
          "typicalItems": 1,
          "usdPerRun": "0.0002",
          "usdPerItem": "0",
          "laterRunShare": 1,
          "basis": "Open-Meteo answers without a key or fee; a run makes one request per city and needs well under a second of CPU"
        }
      }
    ],
    "processing": [
      "weather: unit_convert, threshold_flag (strictly below), dedupe by observation_id (incremental)",
      "documents: files.sync (Google Drive, Confluence or HTTP file listing; only new or changed revisions are processed)",
      "document.normalize (text, HTML, PDF, OCR of scans and images, transcription of audio recordings)",
      "facts.extract (Claude over a fixed or generated fact schema, gated before storage) or basic.listing (rules, no LLM)",
      "facts.reconcile (a newer revision supersedes the facts of the one it replaced)",
      "per-item quarantine with a reason code",
      "citations: source, revision and the verbatim supporting snippet"
    ],
    "units": [
      "C",
      "F",
      "K"
    ],
    "storage": [
      "file://",
      "postgres://"
    ],
    "outputs": [
      "mcp",
      "docker image (linux/amd64)"
    ],
    "schedule": {
      "timezone": "UTC",
      "forms": [
        "interval: {interval_seconds} from 60 s to 1 year, counted from the previous run",
        "calendar: {cron, timezone: 'UTC'} with a 5-field cron (minute hour day-of-month month day-of-week), e.g. '0 8 * * 1' = Mondays 08:00"
      ],
      "request": "state it in the intent, in hints.schedule or as answers.schedule (seconds, a phrase like 'daily at 06:00' or 'every Monday at 08:00', or a cron); the quote's plan.schedule and schedule.description show what will run",
      "runtime": "the image schedules itself in UTC: on empty storage it starts idle and run_now loads the first data (the first scheduled run follows one slot later), a restart runs once only when a slot was missed, run_now runs any time, describe shows next_run_at; a hosted endpoint gets its first run_now from senga; RUN_INTERVAL_SECONDS=<seconds> is a test shortcut"
    },
    "custom": {
      "sources": "https sources fixed at design time: public ones, and APIs that take a static API key or token in a header or query parameter, optionally on your own origin (e.g. your tenant); you inject the values at docker run, each credential is sent only to its source's host and never reaches senga. OAuth, interactive logins, cookies and request signing are not supported",
      "credentials": "the quote's config_schema declares each variable you inject at docker run (secrets carry only a <placeholder>), its credentials list says which single host each is sent to and how; the manifest repeats them with sample_config. The build test runs with placeholders, so a source that needs your key is verified to answer or fail cleanly with the upstream_auth tool error",
      "tools": "up to 8 MCP tools with typed arguments, plus describe, run_now and run_status",
      "citations": "every data answer carries citations: source_id (a stable hash of the source URL), revision (ETag, Last-Modified or the body hash at download), source_url and a verbatim supporting_snippet; the harness keeps each downloaded source and marks every citation verified or unverified against it, and the build test requires a verified citation from each tool",
      "storage": "persistent key-value store with optional background collection; STORAGE_URI file:///data (default) or postgres:// (the image's own Postgres for a localhost URI, or yours); switching is a config change and a fresh run_now",
      "delivery_seconds": 600
    },
    "documents": {
      "description": "Any task over the buyer's own documents (extract, classify, list, find or track facts) that no predefined pipeline covers: describe it in the intent and the quote designs a fact schema, then assembles known steps (sync, normalize with OCR or speech-to-text when needed, extraction to that schema, a generic citation gate) into the runtime image. Quoted as a plan (work_kind assemble_known_steps); the schema is frozen in plan.steps[facts.extract].params.definition and shown as extraction_schema.",
      "sources": [
        {
          "type": "google-drive",
          "credentials": [
            {
              "env": "DRIVE_FOLDER_ID",
              "required": true,
              "secret": false,
              "description": "Folder to read: transcripts (Google Docs, .txt, .vtt) and audio-only recordings directly in it.",
              "example": "<ID of the Google Drive folder with the meeting transcripts>"
            },
            {
              "env": "GOOGLE_SA_JSON",
              "required": true,
              "secret": true,
              "description": "Google Cloud service account key (JSON) the folder is shared with as Viewer; used only for the Drive API (drive.readonly).",
              "example": "<service account JSON with read access to the folder>"
            }
          ]
        },
        {
          "type": "confluence",
          "credentials": [
            {
              "env": "CONFLUENCE_BASE_URL",
              "required": true,
              "secret": false,
              "description": "Base URL of the Confluence Cloud site.",
              "example": "<base URL of the Confluence Cloud site, e.g. https://your-site.atlassian.net/wiki>"
            },
            {
              "env": "CONFLUENCE_SPACE_KEY",
              "required": true,
              "secret": false,
              "description": "Space whose pages and attachments are read.",
              "example": "<key of the Confluence space, e.g. ENG>"
            },
            {
              "env": "CONFLUENCE_EMAIL",
              "required": true,
              "secret": false,
              "description": "Atlassian account used with the API token (basic auth).",
              "example": "<e-mail of the Atlassian account that owns the API token>"
            },
            {
              "env": "CONFLUENCE_API_TOKEN",
              "required": true,
              "secret": true,
              "description": "API token of that account; read access to the space's pages and attachments is enough.",
              "example": "<Atlassian API token with read access to the space>"
            }
          ]
        },
        {
          "type": "http-files",
          "credentials": [
            {
              "env": "SOURCE_URL",
              "required": true,
              "secret": false,
              "description": "HTTP(S) file listing of the archive ({ items: [{ id, revision, sha256, mime_type, download_url }] }); downloads honour ETag / If-None-Match.",
              "example": "<http(s):// URL of the archive's file listing>"
            }
          ]
        },
        {
          "type": "sharepoint_drive",
          "credentials": [
            {
              "env": "SHAREPOINT_DRIVE",
              "required": true,
              "secret": false,
              "description": "Document library to read: its URL (a subsite path such as /sites/Team/EU is kept; a site URL without a library reads the site's default library), a site ID (default library), or the ID of a document library drive (b!\u2026).",
              "example": "<URL of the SharePoint document library or site, e.g. https://contoso.sharepoint.com/sites/Team/Shared%20Documents, or the library's drive ID>"
            },
            {
              "env": "SHAREPOINT_FOLDER",
              "required": true,
              "secret": false,
              "description": "Folder to read with all its subfolders: a path inside the document library, or a drive item ID.",
              "example": "<path of the folder in the document library, e.g. /Reports, or / for the whole library>"
            },
            {
              "env": "GRAPH_TENANT_ID",
              "required": true,
              "secret": false,
              "description": "Microsoft Entra tenant of the app registration.",
              "example": "<Microsoft Entra tenant ID (GUID) of the app registration>"
            },
            {
              "env": "GRAPH_CLIENT_ID",
              "required": true,
              "secret": false,
              "description": "Application (client) ID of the app registration; Microsoft Graph application permission Sites.Read.All, or Sites.Selected with read access to the site.",
              "example": "<application (client) ID of the app registration with Sites.Read.All or Sites.Selected>"
            },
            {
              "env": "GRAPH_CLIENT_SECRET",
              "required": true,
              "secret": true,
              "description": "Client secret of the app registration, used for a client-credentials token.",
              "example": "<client secret of the app registration>"
            }
          ]
        },
        {
          "type": "sftp_folder",
          "credentials": [
            {
              "env": "SFTP_HOST",
              "required": true,
              "secret": false,
              "description": "SFTP server to read from.",
              "example": "<host name or IP address of the SFTP server, e.g. sftp.contoso.com>"
            },
            {
              "env": "SFTP_PORT",
              "required": false,
              "secret": false,
              "description": "TCP port of the SFTP server.",
              "example": "22"
            },
            {
              "env": "SFTP_PATH",
              "required": true,
              "secret": false,
              "description": "Folder to read with all its subfolders; a relative path starts in the login directory. Hidden entries (names starting with a dot) and symbolic links are skipped.",
              "example": "<path of the folder on the SFTP server, e.g. /exports/reports>"
            },
            {
              "env": "SFTP_USERNAME",
              "required": true,
              "secret": false,
              "description": "SFTP account to log in with; it needs read access to the folder only.",
              "example": "<user name of the SFTP account>"
            },
            {
              "env": "SFTP_PASSWORD",
              "required": false,
              "secret": true,
              "description": "Password of the SFTP account (also answers a keyboard-interactive prompt). Set this or SFTP_PRIVATE_KEY.",
              "example": "<password of the SFTP account>"
            },
            {
              "env": "SFTP_PRIVATE_KEY",
              "required": false,
              "secret": true,
              "description": "Private key of the SFTP account, the whole key file content without a passphrase (newlines may be written as \\n). Set this or SFTP_PASSWORD.",
              "example": "<unencrypted private key of the SFTP account in OpenSSH or PEM format>"
            },
            {
              "env": "SFTP_HOST_KEY",
              "required": false,
              "secret": false,
              "description": "Recommended: the expected host key fingerprint (SHA256, base64, as ssh-keygen -lf prints it), so the password or key is only sent to the real server; the run fails with SFTP_HOST_KEY_MISMATCH when the server presents another key. Without it any host key is accepted and every start logs a sftp_host_key_unpinned warning with the presented fingerprint to copy here.",
              "example": "<SHA256 fingerprint of the server's host key as ssh-keygen -lf shows it, e.g. SHA256:abc\u2026>"
            }
          ]
        }
      ],
      "schema": {
        "limits": {
          "record_types": 4,
          "fields_per_type": 8,
          "field_types": [
            "text",
            "number",
            "date",
            "enum",
            "boolean"
          ],
          "enum_values": 12
        },
        "json_schema": "components.schemas.GeneratedSchema in /openapi.json",
        "example": {
          "record_types": [
            {
              "name": "invoice",
              "description": "An invoice the document states",
              "fields": [
                {
                  "name": "number",
                  "type": "text",
                  "description": "Invoice number",
                  "required": true
                },
                {
                  "name": "total",
                  "type": "number",
                  "description": "Total amount",
                  "required": false
                },
                {
                  "name": "due_date",
                  "type": "date",
                  "description": "Payment due date",
                  "required": false
                },
                {
                  "name": "status",
                  "type": "enum",
                  "values": [
                    "paid",
                    "open"
                  ],
                  "description": "Payment status",
                  "required": false
                },
                {
                  "name": "disputed",
                  "type": "boolean",
                  "description": "Whether it is disputed",
                  "required": false
                }
              ]
            }
          ]
        },
        "guarantee": "generated schema: records are guaranteed to pass the generic gate (cited segment sent and quoted verbatim, type in the schema, required fields filled, enum values listed, numbers numeric, ISO dates stated in the document); extraction accuracy is not measured for it"
      },
      "tools": [
        "fact_types",
        "search_facts",
        "facts_by_type",
        "fact_conflicts",
        "search_documents"
      ],
      "validation_gate": [
        "an item whose revision and sha256 are unchanged is skipped (UNCHANGED) without downloading it again; content with identical bytes or the same normalized text is extracted once and its other copies become alias citations (DUPLICATE_CONTENT)",
        "an empty file is quarantined (EMPTY_SOURCE), an unreadable one (CORRUPT_DOCUMENT, UNSUPPORTED_FORMAT, OCR_EMPTY) too, while the rest continue",
        "a record is kept only when its type is in the schema and it cites a segment that exists and was sent to extraction; its supporting_snippet is that segment verbatim",
        "a record whose required field is empty is dropped; an enum value outside its listed values, a number with no numeric value or a date that is not an ISO date (YYYY-MM-DD) stated in the document is dropped from a required field and set to null in an optional one",
        "a missing value stays null, never invented; identical facts from duplicate content are answered once with every citation",
        "a new revision replaces the facts of its source; every answer cites source_id, revision, locator and the snippet; each dropped record appears with its code and reason in the validate log line of its content, under the first item with that content, and run_status counts the drops by code (extract_dropped)"
      ],
      "storage": [
        "file://",
        "postgres://"
      ],
      "extraction": "Optional senga fallback. Only plans that extract facts get this fallback. With the buyer's own ANTHROPIC_API_KEY the image never calls senga; the order key stays in plan.json unused. Without it, the image calls senga's POST /v1/extract with that key. The key is a metered extraction credential for this order only, not a buyer secret and not a key to the model provider: it can spend only this order's extraction budget (the quote's extraction.budget_usd), and senga stops accepting calls once that budget is spent. Facts and normalized content stay in the buyer's storage; a call sends senga only the documents' text segments with their titles, dates and speakers.",
      "delivery_seconds": 900
    }
  },
  "scenarios": {
    "failing_build": {
      "purpose": "A deterministic order whose built image fails its test, to check that a failed build settles nothing. No model, router or geocoding is involved.",
      "quote": {
        "method": "POST",
        "url": "https://senga.ngrok.app/v1/quotes",
        "body": {
          "intent": "Published failing build: build an image that fails its test, to check that nothing is charged.",
          "hints": {
            "scenario": "failing_build"
          }
        }
      },
      "quoted": "answers quoted at once with scenario failing_build, the regular price and a weather plan whose mcp_tools include unserved_tool, a tool the runtime does not serve",
      "order": "pay it like any quote: POST /v1/orders {quote_id} \u2192 402, resend with PAYMENT-SIGNATURE \u2192 202 with order_id, access_token and status_url",
      "expect": {
        "phases": [
          "accepted",
          "building",
          "testing",
          "failed"
        ],
        "status": "failed",
        "error_code": "test_failed",
        "test_report": "passed false; the healthz check fails because the image refuses to start with a tool it does not serve",
        "settlement": "payment.settlement stays null: settle is never called and the facilitator moves no money",
        "authorization": "the signed authorization is never settled and lapses at its validBefore",
        "artifact": "artifact_url answers 409 not_ready; there is no image to download",
        "replay": "replaying the same PAYMENT-SIGNATURE answers 200 with the same failed order, never a charge"
      }
    },
    "topup_required": {
      "purpose": "A deterministic underpriced order: the quoted price covers one generation attempt, the first attempt fails its test on purpose, so the order pauses for a top-up. No model, router or geocoding is involved; it works whatever GENERATION_ATTEMPTS_INCLUDED is.",
      "quote": {
        "method": "POST",
        "url": "https://senga.ngrok.app/v1/quotes",
        "body": {
          "intent": "Published underpriced order: a generated server whose first attempt fails its test, to check the mid-order price increase.",
          "hints": {
            "scenario": "topup_required"
          }
        }
      },
      "quoted": "answers quoted at once with scenario topup_required, the custom server price and a spec with the single tool scenario_status",
      "order": "pay it like any quote: POST /v1/orders {quote_id} \u2192 402, resend with PAYMENT-SIGNATURE \u2192 202 with order_id, access_token and status_url",
      "expect": {
        "phases": [
          "accepted",
          "building",
          "testing",
          "topup_required"
        ],
        "status": "topup_required",
        "topup": "the status answer carries topup: extra_usd, total_usd, new_delivery_seconds, reason, expires_at, pay and cancel",
        "pay": "POST topup.pay.url (or the MCP tool pay_topup) \u2192 402 for topup.total_usd; resend with a fresh PAYMENT-SIGNATURE \u2192 202; the second attempt passes its test, settles the new total once and the order is delivered",
        "cancel": "or POST topup.cancel.url (MCP cancel_order) \u2192 the order fails with cancelled; nothing is charged",
        "expiry": "an unanswered offer expires at topup.expires_at and the order fails with topup_declined; nothing is charged"
      }
    }
  },
  "order_states": [
    "accepted",
    "building",
    "testing",
    "ready",
    "payment_required",
    "topup_required",
    "settling",
    "delivered",
    "failed"
  ],
  "pricing": [
    {
      "product": "meetings",
      "amount_usd": "0.001",
      "asset": "USDC",
      "network": "eip155:8453",
      "networks": [
        "eip155:8453",
        "eip155:84532"
      ],
      "per_run_cost_usd": "0.19 at typical usage; see the quote's run_cost_estimate and platform.pipelines"
    },
    {
      "product": "wiki",
      "amount_usd": "0.001",
      "asset": "USDC",
      "network": "eip155:8453",
      "networks": [
        "eip155:8453",
        "eip155:84532"
      ],
      "per_run_cost_usd": "0.75 at typical usage; see the quote's run_cost_estimate and platform.pipelines"
    },
    {
      "product": "http_archive",
      "amount_usd": "0.001",
      "asset": "USDC",
      "network": "eip155:8453",
      "networks": [
        "eip155:8453",
        "eip155:84532"
      ],
      "per_run_cost_usd": "0.01 at typical usage; see the quote's run_cost_estimate and platform.pipelines"
    },
    {
      "product": "weather",
      "amount_usd": "0.001",
      "asset": "USDC",
      "network": "eip155:8453",
      "networks": [
        "eip155:8453",
        "eip155:84532"
      ],
      "per_run_cost_usd": "0.00 at typical usage; see the quote's run_cost_estimate and platform.pipelines"
    },
    {
      "product": "custom_mcp_server",
      "amount_usd": "0.01",
      "asset": "USDC",
      "network": "eip155:8453",
      "networks": [
        "eip155:8453",
        "eip155:84532"
      ],
      "per_run_cost_usd": "0.00"
    },
    {
      "product": "documents_pipeline",
      "amount_usd": "0.001",
      "asset": "USDC",
      "network": "eip155:8453",
      "networks": [
        "eip155:8453",
        "eip155:84532"
      ],
      "per_run_cost_usd": "see the quote's run_cost_estimate"
    }
  ],
  "payment": {
    "protocol": "x402",
    "version": 2,
    "scheme": "exact",
    "asset": "USDC",
    "network": "eip155:8453",
    "networks": [
      "eip155:8453",
      "eip155:84532"
    ],
    "pay_to": "0x90159e4f1aa11AB09d0362ea6ee32B1fd870cB0c",
    "facilitator": "https://api.cdp.coinbase.com/platform/v2/x402",
    "max_timeout_seconds": 300
  },
  "endpoints": {
    "openapi": "https://senga.ngrok.app/openapi.json",
    "quotes": "https://senga.ngrok.app/v1/quotes",
    "orders": "https://senga.ngrok.app/v1/orders",
    "order_status": "https://senga.ngrok.app/v1/orders/%7Border_id%7D",
    "artifact": "https://senga.ngrok.app/v1/orders/%7Border_id%7D/artifact",
    "image": "https://senga.ngrok.app/v1/orders/%7Border_id%7D/artifact/image.tar",
    "mcp": "https://senga.ngrok.app/o/%7Border_id%7D/mcp",
    "order_payment": "https://senga.ngrok.app/v1/orders/%7Border_id%7D/payment",
    "recover": "https://senga.ngrok.app/v1/orders/%7Border_id%7D/recover",
    "recover_by_body": "https://senga.ngrok.app/v1/orders/recover",
    "a2a": "https://senga.ngrok.app/a2a",
    "platform_mcp": "https://senga.ngrok.app/mcp",
    "order_events": "https://senga.ngrok.app/v1/orders/%7Border_id%7D/events",
    "registry": "https://senga.ngrok.app/v2/"
  },
  "errors": {
    "intent_required": "POST /v1/quotes needs intent (or quote_id + answers)",
    "invalid_delivery": "POST /v1/orders delivery is optional; when sent it must be self_hosted (the default: pull or download the image and run it yourself) or hosted",
    "hosting_needs_buyer_secrets": "delivery hosted is refused for a plan that needs the buyer's secrets; order it without delivery (self_hosted) and run the image yourself; nothing was charged",
    "quote_required": "POST /v1/orders needs {quote_id} from the free POST /v1/quotes (q_example only shows the price)",
    "quote_not_found": "unknown quote_id",
    "quote_expired": "quote passed expires_at; request a new one",
    "quote_not_priced": "quote still needs answers or is unsupported",
    "quote_final": "quote is priced or unsupported and can no longer change, except run_cost_estimate inputs and the source values in source_setup.values; order it or start a new one",
    "quote_agent_unavailable": "the quote or design agent, or geocoding, is temporarily unavailable; retry shortly",
    "input_too_large": "intent, hints and answers of one quote are too long together",
    "body_too_large": "the request body exceeds the size limit",
    "rate_limited": "too many quote, order or recovery requests; retry after the Retry-After header",
    "payment_required": "no PAYMENT-SIGNATURE header (pay the 402 challenge), or the authorization expired before settlement (order status payment_required: POST /v1/orders again with the same quote_id and a fresh PAYMENT-SIGNATURE to pay for the tested image)",
    "payment_in_progress": "a payment for this waiting order is being verified; poll its status_url",
    "unpaid_orders": "this payer has the maximum number of tested orders waiting in payment_required; pay one first; nothing was charged",
    "invalid_payment": "PAYMENT-SIGNATURE could not be decoded, was rejected by the facilitator, does not match the quoted price, or belongs to another order or quote",
    "capacity_exhausted": "no build capacity right now; nothing was charged",
    "hosting_full": "delivery hosted is refused before any payment while every hosted MCP endpoint slot is in use; omit delivery (self_hosted) to order now, or retry hosted after retry_after_seconds (also the Retry-After header); nothing was charged. On a delivered order's hosting.error it means the slot was gone at delivery: run the image yourself",
    "too_many_orders": "this payer already has the maximum number of orders in progress; nothing was charged",
    "unauthorized": "no access_token: use the order's URLs as returned (?access_token=), X-Access-Token or Authorization: Bearer",
    "forbidden": "the access_token does not belong to this order (or was revoked by a wallet recovery)",
    "invalid_order_id": "order_id is not of the form ord_\u2026",
    "invalid_recovery": "POST /v1/orders/{order_id}/recover takes {} for a challenge or {challenge_id, signature} to redeem it",
    "challenge_unknown": "no such recovery challenge; request a new one",
    "challenge_used": "the recovery challenge was already used; request a new one",
    "challenge_expired": "the recovery challenge expired; request a new one",
    "challenge_mismatch": "the recovery challenge was issued for another order",
    "wallet_mismatch": "the recovery challenge was not signed (EIP-191) by the wallet that paid the order; the challenge is spent",
    "recovery_busy": "too many open recovery challenges; retry shortly",
    "quote_already_ordered": "this quote already has a live order (order_id in the answer); follow that order or request a new quote. Nothing was charged",
    "not_waiting_for_payment": "POST /v1/orders/{order_id}/payment takes a payment only for an order in payment_required (or one that failed); this order is past that (order_status in the answer), follow it at its status_url. Nothing was charged",
    "payment_window_closed": "the waiting order's quote expired before a fresh authorization arrived; the tested image was removed and nothing was charged",
    "not_waiting_for_topup": "only an order in topup_required takes a top-up payment or a cancel",
    "topup_expired": "the top-up offer expired; nothing was charged",
    "topup_declined": "the top-up offer expired without payment; the order failed and nothing was charged",
    "cancelled": "the buyer cancelled the paused order; nothing was charged",
    "access_recovered": "the paying wallet recovered this order's access; a replayed payment no longer returns a token, use the recovered one",
    "order_not_found": "unknown order_id",
    "not_ready": "the order is not delivered yet; poll its status_url",
    "build_failed": "image build failed (or the server could not be generated); nothing was charged",
    "test_failed": "built image failed its test (a generated server after every attempt, or the published failing_build scenario); nothing was charged",
    "settlement_failed": "settlement failed after a passing test; nothing was charged",
    "interrupted": "the platform restarted before the order completed; nothing was charged",
    "settlement_unknown": "settlement was attempted but its outcome is unknown; the tested artifact is kept for review; check the payment transaction before buying again",
    "hosting_failed": "the artifact is delivered, but its hosted MCP endpoint could not be started",
    "hosting_expired": "the hosted MCP endpoint ran its time (hosting.expires_at, mcp.expires_at) and was stopped; hosting.status is expired with expired_at. The delivered image stays available: pull or download it from the artifact and run manifest.run",
    "internal_error": "unexpected platform error; nothing was charged",
    "platform_unavailable": "502 or 503 from the proxy in front of senga while senga restarts or deploys; the body may be the proxy's HTML page without a code. Transient: retry the same request after a few seconds (Retry-After when present); nothing changed on your order",
    "not_found": "no such route; see /.well-known/agent-card.json",
    "too_many_streams": "too many open order event streams (5 per client, 100 in total); poll status_url instead",
    "access_revoked": "an order event stream ended because the paying wallet recovered the order's access; reconnect with the new access_token",
    "invalid_request": "POST /v1/extract body does not match {order_id, schema, document, segments} or {order_id, schema, documents: [{document, segments}]} with distinct documents",
    "too_many_segments": "POST /v1/extract takes at most 400 segments per request, across all documents of a batch; split the content",
    "invalid_extract_key": "POST /v1/extract needs Authorization: Bearer <the order's extract key from plan.json>",
    "not_delivered": "POST /v1/extract opens once the order is settled",
    "extract_busy": "POST /v1/extract takes at most 3 requests per order at a time; retry after a running one answers",
    "extract_budget_exhausted": "the order's extraction budget is spent; normalized content stays searchable, new facts are not extracted",
    "extract_daily_limit": "senga's extraction spend for the UTC day is used up; retry after 00:00 UTC, normalized content stays searchable",
    "llm_failed": "the extraction model failed; nothing was charged, retry in the next run",
    "unknown_schema": "POST /v1/extract schema is neither a registered schema nor order:<order_id> of an order whose plan froze a generated schema",
    "schema_mismatch": "POST /v1/extract schema order:<order_id> must name the order of the request"
  }
}