codebase-oracle

Semantic search across all your local repos, via MCP or CLI.

codebase-oracle builds one semantic index over every git repo under a root directory, then exposes it to agents via MCP or to humans via CLI. The vector store lives on your machine; embeddings are computed by OpenAI by default, or fully local via Ollama (configurable). Indexing is incremental: only new and changed files are re-embedded. Built for agents first, humans second.

Try it in 60 seconds

git clone https://github.com/LanNguyenSi/codebase-oracle.git
cd codebase-oracle
npm install && npm run build

# point at the directory holding your git repos, set your key
export ORACLE_SCAN_ROOT=~/code
export OPENAI_API_KEY=sk-...

# build the index, then ask a question
npm run index
npm run query -- "where do we handle auth?"

Or wire it into Claude Code as an MCP server:

claude mcp add codebase-oracle -- npx tsx src/mcp-server.ts

From any Claude Code session on the same machine you can now call oracle_search, oracle_query, oracle_expand, and oracle_list_repos against the shared index.

What a search looks like

oracle_search with query="where do we read AGENT_TASKS_TOKEN" returns matching chunks with line-number locations:

[1] src/auth/token.ts:14-32 (agent-tasks-cli):
function loadToken(): string {
  const value = process.env.AGENT_TASKS_TOKEN;
  if (!value) throw new Error("AGENT_TASKS_TOKEN missing");
  return value;
}

---

[2] backend/src/middleware/auth.ts:8-21 (agent-tasks):
export function requireToken(req, res, next) {
  const token = req.headers.authorization?.replace(/^Bearer /, "");
  if (token !== process.env.AGENT_TASKS_TOKEN) return res.sendStatus(401);
  next();
}

oracle_list_repos shows what's indexed and how fresh each repo is:

- agent-tasks — 1842 chunks across 287 files (indexed 2026-04-27T10:14:02Z, 14 min ago)
- agent-tasks-cli — 421 chunks across 68 files (indexed 2026-04-27T10:14:18Z, 14 min ago)

Next steps

If you want to...	Read
Wire it into Claude Code (MCP setup, the four tools, HTTP MCP auth)	docs/mcp.md
Switch to Ollama, change embedding models, customise scan filters	docs/configuration.md
Understand how the index is built (chunking, embeddings, sqlite-vec)	docs/architecture.md
Migrate from v0.2 (JSONL) or pick up v0.4 line numbers	docs/upgrades.md

CLI reference

The CLI auto-loads .env from the repo root if present.

npm run index                            # build/refresh the index over ORACLE_SCAN_ROOT
npm run index -- --path /path/to/repos   # custom scan root
npm run query -- "what is the audit system?"
npm run query -- -r my-repo "where is the schema defined?"
npm run query -- -k 20 "list all API endpoints"
npm run dev -- search "evaluateTransitionRules"
npm run watch                            # keep the index fresh in the background

Flag	Description
-r, --repo <name>	Filter results to a specific repo
-k, --limit <n>	Number of chunks to retrieve (default: 12)

Watch mode runs a chokidar watcher over the scan root and re-embeds changed files after a short debounce. Newly dropped .git roots need one explicit npm run index to back-fill before watch mode picks up subsequent edits. See docs/architecture.md for details.

Two use cases

For agents (primary). A local Claude Code or other MCP client talks to the oracle's MCP server over stdio. The agent runs oracle_search / oracle_query / oracle_expand / oracle_list_repos against a shared, pre-built index: it never has to scan the filesystem, embed anything, or burn its own context on grep output. One scan for everyone, semantic instead of regex, no duplicate embeddings, MCP-first design.

For humans. The CLI is useful for spot checks, debugging the index, or terminal-driven answers without going through an agent.

Development

npm run build          # TypeScript compilation
npm test               # vitest run
npx tsc --noEmit       # type check only

License

MIT. See docs/architecture.md#credits for inspiration and prior art.