Documentation

Quickstart

From zero to your first code search in under ten minutes, using Claude Code.

This is the fast path to a working setup in Claude Code. About ten minutes, four steps.

There is no separate indexing step. With a PROMETHEUS_API_KEY the context server runs in managed mode — on first start it indexes your workspace into a local SQLite database (~/.prometheus/<hash>.db), embeds the chunks via the prom.codes proxy, and keeps watching for file changes. Your code never leaves your machine; only embedding text transits to the proxy.

1. Get an API key

Sign in at app.prom.codes/app/api-keys → mint a prom_live_… key. This one key unlocks managed code embeddings and memory. Copy it; you'll paste it in the next step.

2. Add the servers

Run these in your project directory. The -- separator before the command is required.

claude mcp add context --env PROMETHEUS_API_KEY=prom_live_… -- npx -y @prom.codes/context-mcp@latest
claude mcp add memory  --env PROMETHEUS_API_KEY=prom_live_… -- npx -y @prom.codes/memory-mcp@latest
claude mcp add saver -- npx -y @prom.codes/saver@latest

context is code search, memory is durable agent memory, and saver (optional, no key needed) trims token cost. The workspace is auto-detected — no PROMETHEUS_WORKSPACE_ROOT needed.

After adding memory and saver, ask your agent to run memory_setup and setup once per workspace — they install the memory-protocol and efficient-output rules into your runtime configs. Details in Agent Memory and Token Saver.

3. Verify

claude mcp list

You should see context, memory, and saver. Inside a session, run /mcp to confirm the tools are live, or claude mcp get context for one server's details.

For a full end-to-end proof — including that semantic (vector) search is actually live and not just keyword fallback — run the copy-paste self-check. It exercises index_status, a concept search with vector provenance, and the memory round-trip in one go.

4. Ask a real question

Try one of these in a Claude Code session:

  • “Find the function that validates incoming webhook payloads in this workspace.”
  • “Show me all callers of chargeCustomer.”
  • “What does RetryQueue import from ./types?”

Claude picks the matching MCP tool (search_code, find_references, expand_context, …) and quotes the result. Run /mcp any time to inspect tool calls when results look off.

What's next