lore

Ask your codebase a question. Get an answer with citations, in seconds.

lore indexes your repos, docs and agent sessions once,
so every question starts from the right code instead of a blank grep.

At a glance

The same questions about a private 1,816-file service, asked of lore and of Claude Code exploring the repo on disk. Same model (claude-opus-5-5), same code.

Question	lore fast	lore deep	Claude Code
"What's included in our entry-level billing plan?"	4.9 s	13.3 s	28.0 s
"How does auth work in our control plane?"	7.0 s	28.1 s	22.3 s

Claude Code gets faster with lore too: connected over MCP, it answered five questions 20% faster than on its own, and up to 29% faster on the broadest one.

More questions, scores and setup: benchmarks.

See it

The auth question from the table, answered by fast mode in 7 seconds. The repos are private, so their names are changed.

$ lore ask "how does auth work in control-plane ?"

The control plane (cp-server) has exactly two security schemes and no API-key plane:

- Auth0 bearer JWT (bearerAuth) covers the whole user-facing surface:
  identity, organizations, members, invitations and regions [6].
- Internal service token (serviceToken) is the shared INTERNAL_SERVICE_TOKEN
  used between the control plane and the regional servers. Operations that
  list it admit nothing else, and no other operation admits it [6].
- No apiKeyAuth. API keys are deliberately excluded, so a leaked workspace key
  cannot reach tenancy management. A test enforces this [3].
- What a key gets there: a workspace key sent to /v1/me is rejected with 401.
  The MCP server, a second indexed repo, reads that 401 as "not a user" and
  degrades instead of failing [12].
  ...

Sources:
[6] apps/cp-server/internal/api/api.go:16-46
[3] apps/cp-server/internal/api/spec_test.go:111-136
[12] src/acme_mcp/controlplane.py:14-23
...
claude-opus-5-5 · 7000 ms

Why lore

An agent working in a repo on disk finds code the slow way: grep, read a file, grep again, with a model round trip at every step. lore does that work once, when you index. A question then costs a 20 ms search and a single model call.

On real repos that means answers in 5 to 8 seconds, 3 to 7 times sooner than Claude Code exploring the same repo. When you need the full picture, deep mode explores with tools and matched Claude Code on every question we scored.

Two ways to ask

	fast	deep
How	One search, one model call	Claude explores with search, grep, read_file and get_session
Typical time	5 to 8 s	15 to 30 s
Best for	"What is X?", "Where is Y?", standups	"How does X work?", tracing a code path

How it works

flowchart LR
    A["repos, docs,<br/>agent traces"] --> B["chunk<br/>(tree-sitter)"]
    B --> C["embed<br/>(bge-small, local CPU)"]
    C --> D[("Postgres<br/>full-text + pgvector")]
    Q["your question"] --> D
    D --> E["best passages"]
    E --> F["Claude writes the answer<br/>with citations"]

Loading

Code is split into functions and classes, embedded on your machine, and stored in Postgres. A question runs a keyword search and a meaning search together and merges the results. Claude only ever sees the passages lore picked, never the whole repo. The full tour is in how lore works.

Quick start

You need Rust 1.95, Docker and an Anthropic API key.

git clone https://github.com/sudopower/lore && cd lore
make db-up                                   # Postgres with pgvector, in Docker
cargo install --path .                       # the lore CLI and server

export ANTHROPIC_API_KEY=sk-ant-... LORE_ADMIN_TOKEN=dev-admin
lore serve                                   # http://localhost:8080

In a second terminal, create a key, index a repo and ask:

export LORE_URL=http://localhost:8080 LORE_ADMIN_TOKEN=dev-admin
TENANT=$(lore admin tenant create --name me | jq -r .id)
export LORE_KEY=$(lore admin key create --tenant "$TENANT" | jq -r .key)

lore repo push ~/code/app
lore ask "how does auth work?"
lore ask "trace a request from the API to the database" --mode deep

The first start downloads the embedding model (about 130 MB). After that, search runs entirely on your machine.

Use it from anywhere

• The CLI: lore ask, lore search, lore repo push.
• The web page: open http://localhost:8080/, paste your key and ask. Handy in a standup.
• Your agents, over MCP:

  claude mcp add --transport http lore "$LORE_URL/mcp" --header "Authorization: Bearer $LORE_KEY"

  Then tell Claude to use it, with a line in CLAUDE.md such as "Start with lore to find the relevant code." Why.
• Agent traces: send OpenTelemetry GenAI spans to /v1/traces, and ask about past sessions.
• Notes and docs: lore doc put indexes anything that isn't in a repo.

Offline

	Works offline?
Indexing and search	Yes. The embedder runs on your CPU.
fast answers	Yes, with a local model (LORE_LLM=openai and llama-server, for example). By default they use the Anthropic API.
deep answers	No. They need Claude's tool use.

Docs

• How lore works: chunking, embedding, retrieval, and what each part sends to the model
• Using lore: tenants and keys, repos, search, traces, documents, answers, the web page, MCP
• Configuration: every environment variable
• Benchmarks: lore vs Claude Code, and how to run the comparison on your repo
• Self-hosting: a VPS with HTTPS and daily backups
• Development: building, testing, and deploying to kind

License

MIT