Skip to content
Jackdaws

Connect

Add the server to your client

Jackdaws runs as a remote Model Context Protocol server behind OAuth. Adding it takes one URL and one credential. What happens after that depends on your client: whether it can lend the server a model, and whether it authenticates as itself or as a shared connector.

The endpoint

One URL, every client

Placeholder · connector-domain · confirm the public endpoint before launch

https://jackdaws-mcp-565939264190.us-central1.run.app/mcp

The service address is live and OAuth-gated. Swap it for the branded hostname before this page ships, so the published URL survives a redeploy.

A trimmed surface lives at /mcp/ide for editors that cap how many tools they will enable at once. Same credential, same verifiers, a shorter list.

Before you subscribe

Inspect the surface first

Tool and prompt listings stay open. Point any MCP client at the server, authorize, and you can read every tool name, description, and schema without a subscription. Calls are the gated part.

An unsubscribed call returns a payment-required error, not an empty result. A lookup the server cannot complete returns a distinct retryable error instead, so a database blip never reads as a lapsed subscription.

Ways in

Pick the path that matches your client

claude.ai connector

Teams working in the browser

  1. Add the server as a custom connector.
  2. Authorize in the browser. Registration is automatic, so there is no client id to paste.
  3. Listings open immediately; calls need an active subscription.
Model
No sampling. Generative calls return a generation contract your own conversation executes, closed by a deterministic validation tool.
Identity
One credential shared by everyone behind the connector. The ledger counts the connector, not the person.

Claude Code / Desktop

Engineers with a client that lends its model

  1. Add the server to your MCP client configuration.
  2. Authorize in the browser on first call.
  3. Run sampling_probe once to confirm which Layer-2 execution path you land on.
Model
Sampling. The generative agents run on the model already in your session, so the model bill stays yours.
Identity
The connecting credential, or a per-user credential when you have one.

Google Antigravity

The agy CLI and the Antigravity IDE

  1. Keep your subscription's client id and secret in your own secret store.
  2. Run the connect script, or write the server entry into Antigravity's MCP config yourself with an Authorization header.
  3. Reload MCP servers. Tokens are short-lived, so re-run when calls start returning 401.
Model
No sampling. Generative calls return a generation contract for the model driving Antigravity to execute.
Identity
The machine-to-machine credential you connected with — one per user.

CI / custom harness

Pipelines and per-user metering

  1. Ask for a machine-to-machine credential during onboarding.
  2. Hold the client id and secret in your own secret store.
  3. Exchange them for a bearer token and call the server directly.
Model
Sampling when your harness advertises it, the generation contract otherwise.
Identity
One credential per user or per pipeline. This is what makes metering, rate limits, and fair use mean one person.

Two boundaries

What the hosted server cannot do

It cannot see your working tree

The server runs in its own container with no shared mount, so an absolute path from your laptop does not resolve there. Send the skill files inline instead: the verifiers on the remote allowlist accept the content itself and run against a bounded per-call sandbox.

Whole-repository work stays local

Audits that sweep a whole tree, a registry, or your session transcripts need filesystem access the hosted server does not have. Those run inside a consulting engagement, on your machines.

Getting access

Subscriptions are provisioned during onboarding rather than self-serve. Tell us which client your team uses and whether you need per-user credentials, and we will set up the credential, the entitlement, and the first audit together.