Connectors
Connect remote MCP servers so reviews and tasks can check what the code alone can’t show: how a feature is used, what the logs say, the data in a table, or a library’s documentation. Connectors are included in paid plans.
Add a connector
Only a GitHub organization owner, or the owner of a personal account, can manage connectors. Workspace settings → Members marks them GitHub owner or Account owner.
On Free, including after a paid plan ends, reviews and tasks don’t use saved connectors, and you can only remove them.
- Open Workspace settings → Connectors and select Add connector.
- Enter a name, the server’s HTTPS endpoint and the request headers it needs, then select Load tools.
- Choose the tools to allow, the repositories, and where to use the connector.
- Select Save connector. Reviews and tasks started from then on can use it.
| Field | What to enter |
|---|---|
| Name | What the connector holds, such as Production analytics. |
| HTTPS endpoint | The server’s Streamable HTTP address, such as https://mcp.your-service.com/mcp. |
| Request headers | The credentials the server expects, such as Authorization: Bearer <token>. Values are stored encrypted and never shown again. |
| Notes for HeyDeer | Optional: what the connector holds and when it helps. |
When you edit a connector, leave a header’s value blank to keep the saved one. If you change the endpoint, enter every value again.
Endpoint and header requirements
- The endpoint can’t include a query string, a
#fragment, a user name or password, or a port other than 443. - It needs a public domain name, not an IP address or a name ending in
.local,.internal,.test,.invalid,.exampleorlocalhost. - A connector can have up to 10 headers, each value up to 8,192 printable ASCII characters.
- A workspace can have up to 50 connectors.
Tools
Read-only tools lists the tools the server marks as read-only. Every other tool, including one the server doesn’t mark, is under Other tools. When you first load the tools, only the read-only ones are allowed.
| Run | Tools it can call |
|---|---|
| Reviews and Autopilot | Allowed read-only tools |
| Fix with HeyDeer and @heydeer mentions | Every allowed tool, read-only or other |
An agent can call only the allowed tools, and decides when one helps. If the server’s tools change, select Reload tools.
Where connectors are used
A connector applies to all repositories or to the ones you select. A task that works in several repositories gets the connectors of any of them.
By default a connector is used only in private repositories, for changes from people with write access. A review or task leaves it out when:
- the repository is public, or a task can change a public repository;
- the pull request comes from a fork;
- the pull request’s author can’t write to the repository.
Choose Public repositories and forks too only for connectors whose data anyone may see, such as public documentation. The review or task page shows which connectors it left out and why.
Connector calls
A review or task page lists each connector call: the connector, the tool, when it ran, how long it took and whether it failed. HeyDeer doesn’t keep what was sent or returned.
Tool calls and the text they return are part of the work, so they add to its credits. If a review fails after calling a connector, its credits aren’t returned. See How credits work.
Privacy and trust
The agent never sees the header values: HeyDeer’s sandbox adds them to the agent’s requests on the way out.
Requests can include code and task context, which the server’s operator handles under its own policies. Even a read-only tool can return private data, so connect only services you trust, use credentials with no more access than needed, such as a read-only database role, and allow only the tools you need. See Security.
If HeyDeer can’t load a connector’s tools, see Troubleshooting.