Skip to main content

Claude skill

The Convolution Labs skill lets Claude drive your account directly from a terminal session — creating and editing agents, publishing them, running them, and composing data pipelines — by calling the same public REST API documented on this site. It doesn't do anything you couldn't do yourself with curl; it just knows the routes, the request shapes, and the platform's rules (like which resources are read-only) so you don't have to look them up.

Install​

curl -sL https://<host>/api/v1/skill | tar xz -C ~/.claude/skills

Replace <host> with your Convolution Labs deployment's domain. This downloads a tarball served directly by the platform and extracts it into Claude's skills directory.

:::note The download always matches your deployment The skill isn't a separately versioned package — it's served by the same running deployment as the API it wraps. Re-running the install command after a platform update re-downloads whatever the skill looks like today, so it can never drift out of sync with the API it's describing. If you're ever unsure whether your local copy is current, re-run the command; there's no version to track. :::

Set up your API key​

  1. In the dashboard, go to Settings → API keys and create a key. Grant it whichever scopes you want the skill to use — read, write, run — the skill respects the scopes your key has and will surface a clear error if it tries something the key isn't allowed to do.

  2. Export it in your shell:

    export CONVOLUTION_API_KEY=cl_xxxxxxxxxxxxxxxx

The skill reads this environment variable to authenticate. It never asks you to paste the key into a prompt, and it will not print it back to you if asked.

What it can and can't do​

The skill can do anything the REST API can do: list, create, update, delete, and publish agents; run agents and poll for results; list data sources, destinations, and connections; create and update connections; trigger syncs.

It cannot create data sources, data destinations, or LLM keys — the API itself doesn't expose those operations, on purpose. A source or destination requires a database connection string, and letting an API key (or an agent acting through one) introduce a new credential or point a sync at an arbitrary destination would be a way to smuggle data out through a connection nobody reviewed. Create your sources and destinations in the web UI first (Dashboard → Data); once they exist, the skill can list them by id and name and wire them into connections on your behalf.

Worked examples​

Create and run an agent:

Create an agent called "Weekly digest" that summarizes new arXiv papers,
then run it with input {"topic": "reinforcement learning"}.

The skill creates the agent via POST /api/v1/agents, and runs it via POST /api/v1/agents/{id}/run, then polls GET /api/v1/runs/{runId} for you until it completes.

Wire up a pipeline:

List my data sources and destinations, then create a connection from
"Production Postgres" to "Warehouse" syncing the orders table hourly.

The skill lists both via GET /api/v1/data/sources and GET /api/v1/data/destinations, resolves the names you gave to ids, and creates the connection via POST /api/v1/data/connections — it cannot create the source or destination themselves, so this only works once both already exist in the dashboard.

Check on a long-running agent:

Is my "Research assistant" agent's last run still going, and what did it
find so far?

The skill looks up recent runs via GET /api/v1/agents/{id}/runs, checks status, and reports back.

Limits that apply to the skill too​

The skill authenticates with your API key, so it's bound by the same rules documented in the REST API reference: the scopes your key was granted, the same 404-for-both-missing-and-not-yours behavior on other people's resources, and the same rate limits (600 reads/hour, 60 writes+runs/hour combined, per key). If the skill hits a rate_limited or insufficient_credits response, it will tell you rather than retry silently.

See also​

  • REST API reference — every route, request and response shape, and error code the skill (and you) can use directly.
  • Run an agent via API — a narrower walkthrough of the run/poll cycle, useful if you're calling the API yourself rather than through the skill.