Connect an AI coding agent (MCP)

Tarkflo runs an MCP server at the same address as the REST API, so an AI coding agent building your integration can read the contract and make live calls without a separate install.

https://api.tarkflo.com/mcp

It takes the same API key as the REST API (see Authentication), sent the same way. There is nothing else to install: no package, no local process, no second key.

Add it to your agent

Claude Code

claude mcp add --transport http tarkflo https://api.tarkflo.com/mcp \
  --header "Authorization: Bearer $TARKFLO_API_KEY"

Claude Desktop

Settings → Connectors → Add custom connector, or add this to your config file directly:

{
  "mcpServers": {
    "tarkflo": {
      "url": "https://api.tarkflo.com/mcp",
      "headers": { "Authorization": "Bearer tf_live_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx" }
    }
  }
}

Cursor

Settings → MCP → Add new MCP server, same JSON shape as Claude Desktop's config above.

As with any API key, read it from an environment variable rather than pasting it into a config file you might commit.

What it can do

Eight tools, each calling the exact same code the matching REST endpoint does:

Tool	Same as	Needs
list_jobs	List jobs	jobs:view
get_job	Get a job	jobs:view
get_job_stats	Get a job's counts	jobs:view
list_workflows	List workflows	templates:view
get_workflow	Get a workflow	templates:view
list_templates	List templates	templates:view
get_template	Get a template	templates:view
submit_candidates	Submit candidates	candidates:create

Plus one resource, tarkflo://openapi.yaml: the full OpenAPI contract, every endpoint, field and error code, for the agent to read instead of guessing a shape.

CV upload stays a REST call. Extract candidate details takes multipart file uploads, and an MCP tool call is JSON — sending CVs as base64 arguments would blow past the API's request size limit after a file or two. An agent building your integration calls that endpoint directly (it's in the openapi resource above); submit_candidates then takes the cv_id it returns.

A call outside a tool's permission answers the same forbidden shape REST does, naming the permission in required — see Authentication for the full scope table. Errors, rate limits and idempotency all work the same as REST, since it's the same server.

Why one tool call can be slower than the equivalent curl

The MCP connection is stateless: every call is its own request, with no session held open between them, so the agent reconnects for each tool call. For a one-off call this is not noticeable; if you're scripting many calls in a loop, REST directly is faster.