Using Prism with an AI agent
Plain-language questions map to Prism tools automatically.
You set Prism up once. After that, ask what you want — the server's instructions teach the agent which tools to call.
How to talk (no tool names)
| You say | Agent should use |
|---|---|
| "What is this repo?" / "Orient me" | repository_dna, landmarks / overview |
| "Is this codebase healthy?" | repository_health |
| "Where does checkout live?" | landmarks / features / find_symbol |
"I'm about to edit src/… — what breaks?" | blast_radius, test_impact |
| "Can I delete this file?" | safe_delete |
| "Review my current changes" | review_changes |
Optional prompts (orient, before_edit, review_diff)
appear in clients that show a prompt picker.
How the server behaves
- Handshake is fast — connect does not index.
- First tool call may take seconds while the index builds; later calls reuse it.
- stdio only — diagnostics on stderr / logging notifications.
- Lists are bounded (
totalCount/truncated). - Paths outside the workspace are refused.
- Every tool is read-only — no consent APIs for agents. See consent and privacy.
Force a specific repo path
Only if auto-detection is wrong:
npx -y @repo-prism/mcp-server --workspace /path/to/repo
# or
PRISM_WORKSPACE=/path/to/repo npx -y @repo-prism/mcp-serverPut the same into the client's args / env if needed.
If something fails
npx -y @repo-prism/cli doctorinside the same project.- Check workspace (
git rootvs cwd). - Confirm Node 26+.
- Disable/re-enable prism in the client, or restart.
- Set
PRISM_WORKSPACEonce and retry.