Zed Agent MCP

Zed’s Agent Panel can call Bifrost through Model Context Protocol. This is separate from Zed editor/LSP support: configure MCP when you want the agent to call analyzer-backed tools such as get_summaries.

Until Bifrost is published through the official MCP registry, configure Bifrost as a custom stdio context server in settings.json.

Configure MCP

Build or install Bifrost first:

cargo install brokk-bifrost --locked --force

For local development, build this checkout and use an absolute path to the debug binary:

cargo build --bin bifrost

Add a context_servers entry to your Zed settings:

{
  "context_servers": {
    "bifrost": {
      "command": "/path/to/bifrost",
      "args": ["--root", "/path/to/project", "--mcp", "symbol|extended"],
      "env": {}
    }
  }
}

For a local checkout build, command should point at /path/to/bifrost/target/debug/bifrost. Always pass an explicit --root; otherwise Bifrost analyzes whatever directory Zed uses as the subprocess working directory.

symbol|extended exposes the analyzer-backed code-intelligence tools used by the Bifrost agent package, including search_symbols, get_summaries, scan_usages_by_location, get_symbol_locations, and related repository discovery tools. Use a smaller or larger MCP toolset only when the host should see a different surface. See MCP Server for the available toolsets.

Open Zed’s Agent settings and confirm the Model Context Protocol section lists the Bifrost server with its tool count. If the server is disabled, enable it there before starting a new Agent thread.

Add Skills

Zed skills are separate from MCP tools. The MCP server makes Bifrost tools available; skills provide reusable agent instructions for when and how to use those tools.

Zed loads skills only from ~/.agents/skills/ and <worktree>/.agents/skills/. It does not discover skills directly from plugins/bifrost-agent/skills. If you want Zed to see the Bifrost code-intelligence skills, use the Bifrost CLI to install them into one of Zed’s supported skill roots:

bifrost --root /path/to/project --install-skills --target project

Use --target global for ~/.agents/skills, or --mode copy when you want a self-contained install instead of a checkout-local symlink. See CLI for the full option list.

The recommended first set is:

• bifrost-code-navigation
• bifrost-code-reading
• bifrost-codebase-search

Workflow skills such as guided review or PR review may depend on host-specific tools and should be installed only after validating that the host provides the needed capabilities.

Validate the Setup

Use a prompt that requires a Bifrost MCP tool result instead of ordinary file reading:

Use the Bifrost MCP get_summaries tool on src/main.rs. Reply with the symbols returned by the tool.

A successful response should name analyzer symbols from the MCP result, such as modules, classes, fields, or functions from the target file. If the agent says it cannot access Bifrost tools and falls back to reading files directly, check that:

• the context_servers entry is present in the active Zed settings file,
• the server is enabled in Agent settings,
• the command path points at an existing Bifrost binary,
• --root points at the project you want analyzed, and
• the thread was created after the server was added or enabled.

Avoid prompts that only ask about README.md or docs files; those can pass through ordinary file reading without proving the MCP server ran.

Can My Agent Run RQL?

The configuration above uses symbol|extended. In a new Agent thread, confirm that the Bifrost tool list includes query_code, then call it with the inline JSON fields {"match":{"kind":"declaration"},"limit":1}. To validate saved RQL, check a workspace file named bifrost-smoke.rql containing (limit 1 (declaration)), then call query_code with {"query_file":"bifrost-smoke.rql"}.

The inline call is canonical JSON. MCP accepts RQL only from a workspace .rql file through query_file. Zed skills remain separate instructions and cannot expose Bifrost tools without the context server. See MCP query and RQL availability for the full surface matrix and Agent Result Safety before making completeness claims.