OpenCode
OpenCode can use Bifrost through a project-local MCP server entry and can load Bifrost’s generic Agent Skills from the same workspace. MCP makes the analyzer tools callable; the skills separately teach OpenCode when and how to use them.
For OpenCode’s underlying host conventions, see the official MCP server and Agent Skills documentation.
Configure MCP
Section titled “Configure MCP”Install Bifrost first:
cargo install brokk-bifrost --locked --forceAdd a project-local opencode.json at the root of the repository you want
Bifrost to analyze:
{ "$schema": "https://opencode.ai/config.json", "mcp": { "bifrost": { "type": "local", "command": [ "/absolute/path/to/bifrost", "--root", "/absolute/path/to/project", "--mcp", "symbol|extended" ], "enabled": true, "timeout": 300000 } }}Use absolute paths for both the Bifrost binary and project root. OpenCode’s default MCP discovery timeout is five seconds; the longer timeout above allows Bifrost to initialize and index a larger workspace on its first connection.
For local Bifrost development, build this checkout and use the debug binary in
the command array:
cargo build --bin bifrostSet the command’s first element to the absolute path to
target/debug/bifrost.
Quit and restart OpenCode after adding or changing MCP configuration. Then verify the server connection from the project root:
opencode mcp listThe output should list bifrost as connected.
Add Skills
Section titled “Add Skills”Install Bifrost’s three generic code-intelligence skills into the project path that OpenCode discovers:
bifrost --root /absolute/path/to/project \ --install-skills --target project --mode copyThis installs:
bifrost-code-navigationbifrost-code-readingbifrost-codebase-search
Restart OpenCode after installing the skills. To inspect the skills OpenCode discovered, run:
opencode debug skillInstalling skills does not start Bifrost or expose analyzer tools by itself. Keep the MCP configuration above enabled when you want OpenCode to call Bifrost.
Validate the Setup
Section titled “Validate the Setup”Start OpenCode from the configured project root and use a prompt that requires both a skill load and a Bifrost MCP result:
Use the bifrost-code-reading skill. Call the Bifrost get_summaries MCP tool on src/main.rs and report only the declared symbol names returned by that tool.Replace src/main.rs with a source file that exists in the target repository.
A successful smoke shows the bifrost-code-reading skill being loaded and a
bifrost_get_summaries tool call before the answer. Avoid prompts that only
ask about README.md or documentation files; those can pass through ordinary
file reading without proving that the analyzer-backed MCP server ran.
Can My Agent Run RQL?
Section titled “Can My Agent Run RQL?”The configuration above uses symbol|extended, so a fresh OpenCode session
should advertise Bifrost’s query_code tool. Ask OpenCode to call it with the
inline canonical JSON fields:
{"match":{"kind":"declaration"},"limit":1}To validate saved RQL, add a workspace file named bifrost-smoke.rql:
(limit 1 (declaration))Then ask OpenCode to call Bifrost query_code with exactly:
{"query_file":"bifrost-smoke.rql"}The inline call is canonical JSON. MCP accepts RQL only from a workspace
.rql file through query_file. See
MCP query and RQL availability for the full
surface matrix and Agent Result Safety before making
completeness claims.