Connect an agent with MCP
The MCP server exposes the same local model used by the CLI. Initialize the target project first:
bunx design-rover init --cwd /absolute/path/to/projectAdd the server to your client
For clients that accept an mcpServers configuration object:
{ "mcpServers": { "design-rover": { "command": "bunx", "args": [ "design-rover@latest", "mcp", "--cwd", "/absolute/path/to/project" ] } }}Replace the path with your project root. Client configuration formats differ; use your client’s stdio-server configuration if it does not use mcpServers. Bun and Node.js 20.19.0 or newer must be available to the process launching the server.
@latest starts the current release. For a reproducible setup, replace it with a release version you have reviewed. Restart the MCP server after changing its version or project path.
Request context for a task
Ask the agent to call get_context before making a UI change, for example:
{ "query": "add a settings form", "path": "src/App.tsx", "maxChars": 12000}Use a path in your project. Inspect the returned freshness, coverage and omission warnings, then read the relevant implementation before reusing an API.
get_context checks supported source observations by default. When saved observations are stale, it returns review/refresh actions with a working directory. It never applies those actions or approves knowledge. Set checkFreshness: false only for an explicitly unchecked saved snapshot. See context retrieval for the full workflow.
Choose the right tool
| Tool | Use |
|---|---|
get_context |
Task/path context within a record and character budget. |
get_component |
An API by ID or unambiguous name; supply source to disambiguate. |
get_system |
Repository freshness and model summary. |
search_system |
Search whole saved knowledge records. |
get_pattern, get_conventions, get_intents, get_exceptions |
Inspect saved design knowledge. |
check_code |
Check a snippet against the saved model. |
explain_violation |
Explain a finding or relevant saved record. |
Pending knowledge is opt-in for tools that support includePending. Missing models and ambiguous names are reported explicitly.
Run watch mode in another terminal to keep observed inventory updated. Subsequent MCP requests read the updated model; an agent’s existing conversation is not refreshed automatically.