MCP Server

Connect AI assistants and agent frameworks to Crowdee via the Model Context Protocol.

The Crowdee API includes a Model Context Protocol (MCP) server. It lets AI coding assistants, agent frameworks, and any MCP-compatible client interact with your projects, pipelines, datasets, and crowdsourcing jobs without writing REST calls manually.

Endpoint

POST https://api.crowdee.ai/v2/mcp

The server uses the Streamable HTTP transport (@hono/mcp). Connect with any MCP client that supports this transport.

Authentication

Authentication is identical to the REST API. Pass your API key on every request:

X-API-Key: crw_YOUR_API_KEY

To target a specific organisation, add:

X-Organization-Id: YOUR_ORG_ID

See Authentication for instructions on generating API keys.

Connecting a client

Most MCP clients accept a server configuration block. Example for a generic HTTP MCP client:

{
  "mcpServers": {
    "crowdee": {
      "type": "http",
      "url": "https://api.crowdee.ai/v2/mcp",
      "headers": {
        "X-API-Key": "crw_YOUR_API_KEY"
      }
    }
  }
}

Verify the connection

curl -X POST https://api.crowdee.ai/v2/mcp \
  -H "X-API-Key: crw_YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","method":"tools/list","id":1}'

A successful response lists all 46 available tools.

Tools

The server exposes 46 tools grouped by domain:

Identity & organisation

Tool	Description
`get_current_user`	Returns the authenticated user and their active organisation.
`list_organizations`	Lists all organisations the current user belongs to.

Projects

Tool	Description
`list_projects`	Lists projects in the active organisation. Accepts `limit` / `offset`.
`get_project`	Returns a single project by `projectId`.
`create_project`	Creates a project. Requires `name`; accepts `description` and `type` (`standard` or `research`).
`update_project`	Updates a project's `name` or `description`.
`list_project_files`	Lists files attached to a project. Accepts `limit` / `offset`.

Datasets

Tool	Description
`list_datasets`	Lists datasets in the active organisation.
`get_dataset`	Returns a dataset with all its versions.
`create_dataset`	Creates a dataset and an initial raw version. Requires `name`, `internalId`, and `modality`.
`delete_dataset`	Deletes a dataset by `datasetId`.

Verification pipelines

Tool	Description
`list_pipeline_catalog`	Returns all 14 built-in verification pipeline definitions including stage breakdowns and file/context requirements.
`run_verification_pipeline`	Starts a verification run on a research project. Requires `projectId`, `pipelineSlug`, and `fileIds`; accepts `contextId` or inline `contextData`. Deducts credits from the organisation balance.
`list_verification_runs`	Lists runs for a project. Accepts `status` filter and `limit` / `offset`.
`get_verification_run`	Returns a run with full stage-by-stage breakdown and associated files.

run_verification_pipeline performs the same credit deduction and context resolution as the REST endpoint POST /v2/projects/:id/verification-runs. The project must be of type research.

Data Platform

Tool	Description
`list_data_platform_pipeline_catalog`	Returns the unified Data Platform pipeline catalog: dataset cleaning, dedup, split, crowd-labeling, and Language Technology pipelines.
`run_data_platform_pipeline`	Runs a pipeline from the catalog. Supports `targetType: "project_file"` (Language Technology pipelines, one run per file) and `targetType: "dataset_version"` for pipelines with `executionMode: "cleaning"` (produces a new derived dataset version) or `"dedup"` (analyses the version in place). Pipelines with `executionMode: "split"` or `"crowd-label"` are not supported via this tool — use the web app or REST API for those.
`list_data_platform_runs`	Lists Data Platform pipeline runs for the active organisation. Accepts `projectId`, `datasetVersionId`, and `status` filters.
`get_data_platform_run`	Returns a Data Platform pipeline run with its per-file, per-stage breakdown.

run_data_platform_pipeline deducts credits the same way as the REST endpoint POST /v2/data-platform/pipelines/:slug/run for the target/execution-mode combinations it supports.

Language Technology pipelines

These tools predate the unified Data Platform catalog above and only cover the LT subset (transcription, OCR, NER, translation, language identification). Prefer list_data_platform_pipeline_catalog / run_data_platform_pipeline for new integrations.

Tool	Description
`list_lt_pipeline_catalog`	Returns all 7 built-in LT pipeline definitions (transcription, OCR, NER, translation, language identification).
`run_lt_pipeline`	Runs an LT pipeline on one or more project files — one run per file. Accepts optional `contextData` (e.g. `{ "target_language": "German" }` for translation).
`get_lt_run`	Returns an LT run record including the structured result payload (transcript, entities, translation, etc.).

AI Platform

Tool	Description
`list_ai_platform_base_models`	Returns the catalog of open-source base models available for finetuning, with size, license, and credit cost.
`create_finetune_dataset`	Assembles a training dataset either from a crowd-labeled Data Platform dataset version (`sourceType: "dataset_version"`, requires a prior `crowd-label-classify` run) or from a Crowd Platform job's accepted answers (`sourceType: "crowd_job"`).
`list_finetune_datasets`	Lists finetune (training) datasets in the active organisation.
`create_finetune_run`	Starts a finetuning run: finetunes a base model on a training dataset. Requires `finetuneDatasetId` and `baseModelSlug`. Deducts the base model's credit cost from the organisation balance.
`list_finetune_runs`	Lists finetune runs, including the resulting model artifact once completed. Accepts a `status` filter.
`get_finetune_run`	Returns a single finetune run by `runId`, including the resulting model artifact once completed.
`list_models`	Lists finetuned model artifacts (name, base model, Ollama tag, status) in the active organisation.

create_finetune_run charges credits the same way as the REST endpoint POST /v2/ai-platform/finetune-runs. Unlike other run-triggering tools, the resulting job isn't dispatched through the queue — the Python finetune-worker polls finetune_jobs directly.

Crowdsourcing

Tool	Description
`list_crowd_jobs`	Lists jobs for a project. Accepts `status` filter and `limit` / `offset`.
`get_crowd_job`	Returns a job with live slot statistics (assigned, submitted, timed-out, answers by status).
`create_crowd_job`	Creates a crowdsourcing job. Requires `projectId`, `name`, `title`, `description`, and `surveyTemplateVersionId`.
`list_crowd_answers`	Lists answers for a job. Accepts `status` filter and `limit` / `offset`.
`accept_crowd_answer`	Accepts a pending answer by `answerId`.
`reject_crowd_answer`	Rejects a pending answer by `answerId` with an optional `reason`.

Content Gathering

Tool	Description
`create_gathering_job`	Creates a content-gathering crowd job: the crowd searches platforms/the web for items matching criteria (keywords, platforms, languages, content types).
`list_gathered_items`	Lists items a crowd has gathered for a content-gathering job. Accepts `reviewStatus` filter and `limit` / `offset`.
`update_gathered_item`	Accepts, rejects, or marks a gathered item as a duplicate by `itemId`.
`gathered_items_to_dataset`	Converts a gathering job's reviewed items (with attached files) into a new dataset.

AI Output Evaluation

Tool	Description
`create_ai_output_evaluation`	Requests a crowd panel to rate the transparency (clarity, evidence, actionability, bias risk) of a completed verification or Language Technology run. Auto-generates the crowd task.
`create_external_ai_output_evaluation`	Requests a crowd panel to rate the transparency of an AI output from your own (non-Crowdee) AI system — pass `modelName`, `verdict`, `explanation` directly. Stores the submission using the input-data tables, then dispatches the same crowd evaluation as `create_ai_output_evaluation`.
`get_ai_output_evaluation_results`	Gets individual crowd ratings and the aggregate transparency score for an evaluation batch.

Continuous Monitoring

Tool	Description
`create_monitoring_schedule`	Creates a recurring schedule that periodically samples a completed run and dispatches an AI output evaluation for it, tracking drift over time.
`list_monitoring_runs`	Lists the tick history for a monitoring schedule, including drift flags and aggregate scores.

Resources

Four MCP resources expose static catalog data:

URI	Description
`crowdee://pipelines/catalog`	All 14 verification pipeline definitions as JSON.
`crowdee://lt-pipelines/catalog`	All 7 LT pipeline definitions as JSON (deprecated alias — see `crowdee://data-platform/pipelines/catalog`).
`crowdee://data-platform/pipelines/catalog`	The unified Data Platform pipeline catalog (cleaning, dedup, split, crowd-labeling, and LT pipelines) as JSON.
`crowdee://ai-platform/base-models/catalog`	The catalog of open-source base models available for AI Platform finetuning as JSON.

Fetch a resource with the standard resources/read call:

curl -X POST https://api.crowdee.ai/v2/mcp \
  -H "X-API-Key: crw_YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","method":"resources/read","params":{"uri":"crowdee://pipelines/catalog"},"id":2}'

Example: run a verification pipeline

# 1. Find a research project
curl -X POST https://api.crowdee.ai/v2/mcp \
  -H "X-API-Key: crw_..." -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","method":"tools/call","params":{"name":"list_projects","arguments":{}},"id":1}'

# 2. Start a verification run
curl -X POST https://api.crowdee.ai/v2/mcp \
  -H "X-API-Key: crw_..." -H "Content-Type: application/json" \
  -d '{
    "jsonrpc":"2.0","method":"tools/call",
    "params":{
      "name":"run_verification_pipeline",
      "arguments":{
        "projectId":"<projectId>",
        "pipelineSlug":"verify-image-metadata",
        "fileIds":["<fileId>"]
      }
    },
    "id":2
  }'

# 3. Poll for completion
curl -X POST https://api.crowdee.ai/v2/mcp \
  -H "X-API-Key: crw_..." -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","method":"tools/call","params":{"name":"get_verification_run","arguments":{"runId":"<runId>"}},"id":3}'

Example: finetune a model on a crowd job's answers

# 1. Assemble a training dataset from a crowd job's accepted answers
curl -X POST https://api.crowdee.ai/v2/mcp \
  -H "X-API-Key: crw_..." -H "Content-Type: application/json" \
  -d '{
    "jsonrpc":"2.0","method":"tools/call",
    "params":{
      "name":"create_finetune_dataset",
      "arguments":{
        "sourceType":"crowd_job",
        "jobId":"<jobId>"
      }
    },
    "id":1
  }'

# 2. Pick a base model, then start the finetuning run
curl -X POST https://api.crowdee.ai/v2/mcp \
  -H "X-API-Key: crw_..." -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","method":"tools/call","params":{"name":"list_ai_platform_base_models","arguments":{}},"id":2}'

curl -X POST https://api.crowdee.ai/v2/mcp \
  -H "X-API-Key: crw_..." -H "Content-Type: application/json" \
  -d '{
    "jsonrpc":"2.0","method":"tools/call",
    "params":{
      "name":"create_finetune_run",
      "arguments":{
        "finetuneDatasetId":"<finetuneDatasetId>",
        "baseModelSlug":"<baseModelSlug>"
      }
    },
    "id":3
  }'

# 3. Poll for completion — resultModelArtifact is set once status is "completed"
curl -X POST https://api.crowdee.ai/v2/mcp \
  -H "X-API-Key: crw_..." -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","method":"tools/call","params":{"name":"get_finetune_run","arguments":{"runId":"<runId>"}},"id":4}'

On this page