punt-quarry

Local semantic search for AI agents and humans.

License CI PyPI Python Working Backwards

Quarry indexes documents in 20+ formats, embeds them with a local ONNX model (snowflake-arctic-embed-m-v1.5, 768-dim), stores vectors in LanceDB, and serves semantic search to Claude Code, Claude Desktop, and the CLI. Everything runs locally — no API keys, no cloud accounts. The embedding model (~120 MB int8) downloads once on first use. CUDA GPUs are auto-detected for faster inference.

Platforms: macOS, Linux

Quick Start

curl -fsSL https://raw.githubusercontent.com/punt-labs/quarry/6f90f11/install.sh | sh

Restart Claude Code, then:

> /ingest report.pdf                    # index a document (runs in background)
> /quarry status                        # after a moment, confirm it's there
> /find "what does the report say about margins"   # search by meaning

Once installed, a plugin hook auto-indexes your current project directory on every session start — you don't need to /ingest your codebase manually.

uv tool install punt-quarry
quarry install
quarry doctor

curl -fsSL https://raw.githubusercontent.com/punt-labs/quarry/6f90f11/install.sh -o install.sh
shasum -a 256 install.sh
cat install.sh
sh install.sh

Remote Server

Run quarry on a GPU server and connect from any Mac or Linux client over TLS.

Server (GPU host, serves remote clients):

export QUARRY_API_KEY=$(openssl rand -hex 32)
curl -fsSL https://raw.githubusercontent.com/punt-labs/quarry/6f90f11/install.sh | sh -s -- --network

Generates TLS certificates, binds daemon to 0.0.0.0, registers a systemd service, and prints a CA fingerprint. NVIDIA GPUs are auto-detected for CUDA inference.

Client (connects to remote server):

curl -fsSL https://raw.githubusercontent.com/punt-labs/quarry/6f90f11/install.sh | sh
quarry login <server-hostname> --api-key <token>

No special flag needed --- the default install runs a local daemon on localhost. quarry login redirects queries to the remote server over wss:// with TOFU certificate pinning.

Claude Desktop

Download punt-quarry.mcpb and double-click to install. Alternatively, quarry install configures Claude Desktop automatically.

Note: Uploaded files in Claude Desktop live in a sandbox that quarry cannot access. Use remember for uploaded content, or provide local file paths to ingest.

Features

  • 20+ formats --- PDFs (with OCR for scanned pages), source code (AST-aware splitting), spreadsheets, presentations, HTML, Markdown, LaTeX, DOCX, images
  • Semantic search --- retrieval is by meaning, not keyword. A query about "margins" finds passages about profitability even if they never use that word
  • Daemon architecture --- one quarryd process loads the embedding model once and serves the CLI and hooks over a versioned REST API (/v1); Claude Code connects through the stdio quarry mcp server
  • Passive knowledge capture --- quarry enable sets up three scoped collections per project: file sync, passive captures (web fetches + session transcripts), and per-agent memory. Captures are separated from the code index so research doesn't pollute code search
  • Named databases --- isolated LanceDB directories with independent sync registries. Switch with use for work/personal separation
  • Research agent --- researcher subagent combines quarry local search with web research, auto-ingests valuable findings

What It Looks Like

Ingest a document

> /ingest report.pdf

▶ Ingesting report.pdf (background)

Check what's indexed

> /quarry

▶ Database: default
  Documents: 47
  Chunks: 1,203
  Size: 12.4 MB
  Model: snowflake-arctic-embed-m-v1.5 (768-dim)

Search by meaning

> /find "what were the Q3 revenue figures"

▶ [report.pdf p.12 | text/.pdf] (similarity: 0.4521)
  Third quarter revenue reached $142M, up 18% year-over-year,
  driven primarily by expansion in the enterprise segment.
  Gross margins improved to 71% from 68% in Q2.

Commands

Slash Commands (Claude Code)

Command What it does
/ingest <source> Ingest a URL, directory, or file
/remember <name> Ingest inline text under a document name
/find <query> Semantic search. Questions get synthesized answers; keywords get raw results
/explain <topic> Search and synthesize an explanation
/source <claim> Find which document a claim comes from
/quarry [sub] Manage: status, sync, collections, databases, registrations

MCP Tools

Tool Purpose Execution
ingest Index a file or URL Background
remember Index inline text Background
register_directory Register directory for sync Background
sync_all_registrations Re-index all registered directories Background
find Semantic search with filters Sync
show Document metadata or page text Sync
list Documents, collections, databases, registrations Sync
status Database statistics Sync
delete Remove document or collection Background
deregister_directory Remove registration (validates; errors if unknown) Sync
use Switch active database Sync

CLI

quarry ingest report.pdf                       # index a file
quarry ingest https://example.com              # index a webpage
echo "notes" | quarry remember --name notes.md # index inline text
quarry find "revenue trends"                   # hybrid search (vector + FTS)
quarry list documents                          # list indexed documents
quarry register ~/Documents/notes              # watch a directory
quarry sync                                    # re-index registered dirs
quarry use work                                # switch database
quarry enable                                  # set up project collections + captures
quarry disable                                 # remove project registration + data
quarry captures init                           # bootstrap the private capture shadow repo
quarry captures push                           # re-scrub + push captures to the shadow
quarry status                                  # database dashboard
quarry doctor                                  # health check
quarry install                                 # set up daemon service, TLS certs, mcp-proxy
quarryd                                        # run the engine in the foreground (normally the service runs it)

# Remote connections
quarry login okinos.local --api-key <token>    # TOFU login to remote server
quarry logout                                  # disconnect, revert to local daemon
quarry remote list --ping                      # show remote config and health

# Agent memory tagging
quarry ingest notes.md --agent-handle claude --memory-type fact
quarry find "deployment steps" --agent-handle claude
echo "key insight" | quarry remember --name insight.md --agent-handle claude \
  --memory-type observation --summary "Key insight from review"

Setup

Quarry works with zero configuration. These environment variables are available for customization:

Variable Default Description
QUARRY_PROVIDER (auto) ONNX execution provider: cpu, cuda, or unset (auto-detect)
QUARRY_API_KEY (none) Bearer token for quarryd (required for a non-loopback bind)
QUARRY_ROOT ~/.punt-labs/quarry/data Base directory for all databases
CHUNK_MAX_CHARS 1800 Max characters per chunk (~450 tokens)
CHUNK_OVERLAP_CHARS 200 Overlap between consecutive chunks

For the full configuration reference, see Architecture section 7.

Passive Knowledge Capture

Beyond explicit /ingest and /find commands, quarry runs as a Claude Code plugin with hooks that capture knowledge automatically during your sessions:

Hook When it fires What it does
Session start On every session start Auto-registers your project directory and syncs it in the background. Your codebase is searchable without manual ingestion.
Web fetch After any WebFetch tool call URLs Claude fetches during research are auto-ingested into the project's <name>-captures collection (or global web-captures if no project is enabled).
Pre-compact Before context compaction Captures the conversation transcript into the project's <name>-captures collection (or global session-notes if no project is enabled).

All hooks are fail-open — failures are ignored and never block Claude Code. Each hook is individually toggleable via .punt-labs/quarry/config.md YAML frontmatter. See AGENTS.md for the full integration model.

Privacy: captures are scrubbed before write

Captures (session transcripts and auto-ingested web fetches) are redacted at write time, before anything touches disk or the <name>-captures/web-captures collection. A single write choke point scrubs, in order, secrets → filesystem paths (/Users//home/<user>/ → ~/) → emails (→ [REDACTED:email]) → the local hostname (→ [REDACTED:hostname]) → profanity. Web-fetch captures also redact the source URL's userinfo, query, and fragment so a URL like …/reset?email=…&token=… cannot leak. Redaction is idempotent and fail-closed (on any scrub error, no capture is written). Deliberately-ingested content (quarry ingest, remember) is not scrubbed — only the automatic capture paths are, since those feed the shared/pushable collections. See DES-036.

Private capture shadow repo (<repo> → <repo>-quarry)

Redacted capture .md files live in the gitignored .punt-labs/quarry/captures/ dir with no durable home. Enable the optional shadow sync to push them to a per-project private repo <repo>-quarry instead. Opt in by uncommenting the shadow: block in .punt-labs/quarry/config.md:

shadow:
  enabled: true                  # off by default — a network + security action
  remote: ""                     # empty → derive <origin>-quarry from the public remote
  acknowledge_unverified: false  # required to push when gh cannot confirm the remote is private

Pre-create the private repo first. quarry does not create it by default (a wrong owner or an accidental public repo would leak). Create <repo>-quarry as private and let the remote derive, or run quarry captures init --create to create it via gh (which verifies visibility is private before any push).

Once enabled, the captures dir becomes a standalone nested git repo. quarry captures push — and the tail of every quarry sync — re-scrubs the staged captures with the same DES-036 scrubber, aborts the commit if any file is not clean, then commits and pushes. Auth reuses your existing git credentials; quarry stores no new secret. The push is fail-open: a network or auth failure never blocks a session.

Visibility is load-bearing. The re-scrub cannot catch every residual (IDN emails, non-/Users//home paths, a hostname from another machine), so those are backstopped only by the remote being private. quarry therefore refuses to push to a verifiably public remote, and requires acknowledge_unverified: true in .punt-labs/quarry/config.md when gh is absent and it cannot confirm the remote is private. quarry doctor reports the shadow state, and flags a required failure if the public repo already tracks capture files (which git rm --cached stops going forward, but an already-pushed capture needs a history purge — git filter-repo/BFG + force-push — coordinated with the repo owner). See DES-039.

How It Works

Quarry runs as a daemon — one quarryd process per machine holds the engine (embedding model, LanceDB, pipeline, sync registry) and serves a versioned REST API under /v1. The CLI and Claude Code hooks reach it over HTTP; Claude Code's MCP tools run through the stdio quarry mcp server, which is itself a thin REST client of the daemon — it loads no engine.

                       HTTP /v1 (TLS, pinned CA)
CLI + hooks  <--------------------------------------->  quarryd
                                                        (one engine)
              stdio, MCP JSON-RPC        HTTP /v1
Claude Code  <-------------------->  quarry mcp  <----->  quarryd

One resident engine means the CLI and hooks reuse the daemon's loaded model instead of a per-invocation ~200 MB reload. Remote and loopback REST connections use TLS with a self-signed, pinned CA — even on localhost.

The target is that every interface — CLI, library, and MCP — is a thin client of that one daemon over a versioned REST contract; this daemon-first model is specified in DES-031 v2 (ACCEPTED, in implementation). MCP is now a pure client: quarry mcp reaches the daemon through QuarryClient (loopback token, or a remote login's TLS + pinned CA), loading no engine of its own. The CLI and library still load the engine in-process on some local calls; those remaining local-engine paths are being removed.

Claude Code's plugin runs quarry mcp directly — no mcp-proxy in quarry's MCP path. quarry install still downloads mcp-proxy (SHA256-verified, correct platform) because it remains a supported tool for other consumers, and configures MCP clients to spawn quarry mcp.

Managing the daemon

quarry install registers quarryd as a per-user background service that starts at login and restarts on crash — launchd on macOS (label com.punt-labs.quarry) and systemd on Linux (unit quarry.service). You rarely need to touch it, but after upgrading the wheel you must restart the service so the new engine loads — a running daemon holds the old code in memory (quarry doctor reports the CLI version, not the daemon's).

macOS (launchd):

launchctl kickstart -k gui/$(id -u)/com.punt-labs.quarry              # restart (use after an upgrade)
launchctl unload -w ~/Library/LaunchAgents/com.punt-labs.quarry.plist  # stop
launchctl load   -w ~/Library/LaunchAgents/com.punt-labs.quarry.plist  # start
launchctl list com.punt-labs.quarry                                    # status (shows PID, or "-" if stopped)

Linux (systemd, user scope):

systemctl --user restart quarry        # restart (use after an upgrade)
systemctl --user disable --now quarry  # stop
systemctl --user enable  --now quarry  # start
systemctl --user status  quarry        # status

quarry doctor confirms quarryd: running and ready. quarry uninstall removes the service and its unit file. To run the engine in the foreground for debugging, run quarryd directly — it blocks until you stop it.

Documentation

Architecture | Z Specification | Design | Agents | Changelog

Development

uv sync                        # install dependencies
make check                     # run all quality gates (lint, type, test)
make test                      # test suite only
make format                    # auto-format code
make docs                      # build LaTeX documents
make eval                      # retrieval-quality eval harness (MRR/success@k, dev/CI only)

Roadmap

Direction tracks the PR/FAQ. Current focus:

  • Retrieval quality — a make eval harness (per-bucket MRR/success@k + a metadata-pollution diagnostic) now baselines search on quarry's own data; embedding levers (contextual embeddings, late chunking) are measured against it before adoption. See docs/eval-harness-design.md.
  • Private capture sync — captures redact PII at write time, and the opt-in per-project private shadow repo (<repo> → <repo>-quarry, above) now pushes the redacted captures off the public repo entirely.

License

MIT