MCP servers

The Model Context Protocol (MCP) lets an AI agent call tools on another service, such as an issue tracker or a documentation search. In AI CodeOps, you connect MCP servers to your workspace and the agents in your sessions can use their tools.

Here AI CodeOps is the client: it connects to MCP servers you add. To use AI CodeOps from your own coding assistant instead, see the AI CodeOps MCP server.

MCP support is in beta.

Remote and local servers

Sessions run on the platform, not on your computer, so they can't start a program on your machine. Use a remote server for anything your agents should call.

Remote servers must:

• be on a public HTTPS address (private networks and localhost are refused);
• speak the MCP Streamable HTTP transport (JSON-RPC over POST, with a JSON or event-stream response);
• accept authentication in static headers, such as a bearer token. Servers that need an OAuth sign-in aren't supported yet, and neither is the older SSE-only transport.

Add a server

1. Open Configure → MCP Servers and choose Add MCP server.
2. Enter a Name, for example "GitHub". AI CodeOps makes the server's ID from it, in lower case with dashes: "GitHub" becomes github, "Docs Search" becomes docs-search. Tool names and policy rules use this ID.
3. Choose Remote (HTTP/SSE URL) and enter the Server URL.
4. Under Headers, add one KEY=value per line, for example:

``text Authorization=Bearer <your token> ``

1. Choose Save MCP server.

The server appears under Servers with Disable and Remove. There is no edit form yet: to change a server, remove it and add it again.

Managing MCP servers needs the MCP-manage permission: owners, admins and DevOps admins by default.

How secrets are stored

Headers (remote) and environment variables (local) are encrypted at rest with AES-256-GCM. They are never sent back to the browser; the list shows only which keys are set, such as "1 header: Authorization". They are decrypted only when a session connects to the server. Server creation, changes and removal are recorded in the audit log.

How agents get the tools

When a session starts, AI CodeOps connects to each enabled remote server and lists its tools. A server that can't be reached is skipped, and the session runs without its tools.

• Tools are named <server>__<tool>, for example github__create_issue.
• A session can offer up to 32 MCP tools in total.
• In each agent step, an agent can make up to 6 rounds of tool calls.
• Each call times out after 30 seconds, and long results are shortened.

You can narrow which tools reach the agents in two ways.

Per agent

On Configure → Agents, a custom agent has a field MCP tools allowed. Enter servers, tools or server__tool names, comma-separated. Leave it blank to allow all.

By workspace policy

A policy rule on the mcp.invoke action matches server/tool. A tool the policy doesn't allow is never offered to the agents. For example, to keep every delete tool on the GitHub server away from your sessions:

{ "id": "no-repo-deletes", "action": "mcp.invoke", "resource": "github/delete_*", "decision": "deny" }

mcp.invoke rules can use allow or deny, not ask. With the None trust posture, MCP tools are denied unless a rule allows them.

Troubleshooting

If a session doesn't use a server's tools:

• Check the server is enabled and is a Remote server.
• Check the URL is public HTTPS and the server supports Streamable HTTP, not only SSE.
• Check the headers: an expired token makes the server refuse the connection.
• Check the custom agent's MCP tools allowed field and your workspace policy.

Next steps

• Agents and workflows
• Policy rules reference
• Dependency audit