io.github.laradji/deadzone

         _                _
      __| | ___  __ _  __| |_______  _ __   ___
     / _` |/ _ \/ _` |/ _` |_  / _ \| '_ \ / _ \
    | (_| |  __/ (_| | (_| |/ / (_) | | | |  __/
     \__,_|\___|\__,_|\__,_/___\___/|_| |_|\___|

    > semantic doc search. local file. no cloud. no key.
    > you ask in english. it answers in snippets.

Status. Vector search wired end-to-end. MCP over stdio. One binary, Linux + macOS, zero telemetry. See releases for the latest tag and the roadmap for in-flight work. The scraper is still the messy half — #64 is honest about it.

The pitch, in one paragraph

Your AI client says "how do I register a tool?". The doc says AddTool. A grep-based index shrugs; a vector index doesn't. Deadzone is the vector index — nomic-embed-text-v1.5 over Turso's native cosine distance, wrapped in a Go binary that speaks MCP over stdio and keeps every byte on your laptop. It is, roughly, Context7 with the internet turned off.

Rules of the deadzone

1. One binary. deadzone. Subcommands for everything. No pip install, no npm i, no docker compose up.
2. The index never leaves. Local Turso file. No account. No API key. No egress on the hot path.
3. Natural language first. Embeddings over cosine. FTS5 is not invited.
4. The binary is the version. The DB is pinned to the binary. Upgrade the binary, the DB follows; don't, and it won't.
5. Fail loudly or not at all. DEADZONE_DB_OFFLINE=1 refuses to guess. Verification failures in the scraper drop the doc, not the run.

Install (pick one; they all converge on the same binary)

# macOS Apple Silicon — the one-liner
brew install laradji/deadzone/deadzone

# Linux — resolve the latest tag once, then pick a flavor
VERSION=$(curl -fsSL https://api.github.com/repos/laradji/deadzone/releases/latest | grep '"tag_name"' | cut -d'"' -f4)
ARCH=amd64    # or arm64

# self-mounting AppImage
curl -L -O "https://github.com/laradji/deadzone/releases/download/${VERSION}/deadzone_${VERSION}_linux_${ARCH}.AppImage"
chmod +x "deadzone_${VERSION}_linux_${ARCH}.AppImage"
mv "deadzone_${VERSION}_linux_${ARCH}.AppImage" deadzone

# or plain tarball (no FUSE needed)
curl -L "https://github.com/laradji/deadzone/releases/download/${VERSION}/deadzone_${VERSION}_linux_${ARCH}.tar.gz" \
  | tar xz --strip-components=1

Both flavors land a deadzone executable in your current directory, so the ./deadzone server snippet below works as-is.

# Container — multi-arch (linux/amd64 + linux/arm64), ships with the DB baked, runs offline by default
docker pull ghcr.io/laradji/deadzone:latest
docker run --rm -i ghcr.io/laradji/deadzone:latest server

The image bakes the binary, libonnxruntime, deadzone.db, and the nomic-embed-text-v1.5 ONNX weights (~230 MB total), and runs as a non-root user out of distroless (no shell, no package manager). DEADZONE_DB_OFFLINE=1 is set in the image so first launch is instant — no download, no volume mount, no --network access required. To refresh the index, pull a newer tag.

Windows is blocked upstream — no libtokenizers.a. Use WSL.

Verify checksums (optional but cheap):

curl -L -O "https://github.com/laradji/deadzone/releases/download/${VERSION}/deadzone_${VERSION}_checksums.txt"
sha256sum  --ignore-missing -c "deadzone_${VERSION}_checksums.txt"   # Linux
shasum -a 256 --ignore-missing -c "deadzone_${VERSION}_checksums.txt"   # macOS

AppImage needs FUSE v2. Most desktops ship it; minimal servers don't. If you get dlopen(): libfuse.so.2, either apt-get install libfuse2 (or dnf install fuse-libs) or pass --appimage-extract-and-run to bypass FUSE entirely.

Run

./deadzone server

That's the quick-start. On first launch it fetches deadzone.db matched to this binary's version, SHA256-verifies, caches it under the platform data dir, and serves. Second launch onwards: zero network. Upgrade the binary and the DB re-fetches on next launch; don't, and the cache is served forever.

MCP client wire-up — native binary (Brew tap, tarball, or AppImage):

{
  "mcpServers": {
    "deadzone": {
      "type": "stdio",
      "command": "/path/to/deadzone",
      "args": ["server"]
    }
  }
}

MCP client wire-up — container (multi-arch on ghcr.io). The image ships with deadzone.db baked, so no volume mount is needed and every container start is offline-instant:

{
  "mcpServers": {
    "deadzone": {
      "type": "stdio",
      "command": "docker",
      "args": ["run", "--rm", "-i", "ghcr.io/laradji/deadzone:latest", "server"]
    }
  }
}

Then, from the client:

search_libraries("terraform aws")                 → ranked (lib_id, version) pairs
search_docs("creating an s3 bucket", lib_id=...)  → snippets, token-budgeted

The two tools

┌─────────────────────────────────────────────────────────────────────┐
│  search_libraries(name, limit?) → []LibraryHit                      │
│  ─────────────────────────────────────────────                      │
│  free text   ──►  vector match against the `libs` table             │
│                   ──► [{lib_id, version, doc_count, match_score}]   │
├─────────────────────────────────────────────────────────────────────┤
│  search_docs(query, lib_id?, version?, tokens?) → []Snippet         │
│  ──────────────────────────────────────────────────                 │
│  natural  ──► 768-dim embed ──► cosine over docs                    │
│  language                      ──► token-budgeted snippets back     │
└─────────────────────────────────────────────────────────────────────┘

Under the hood

  deadzone server
       │
       ▼
  ┌──────────────┐   stdio JSON-RPC       ┌───────────────┐
  │  MCP client  │ ─────────────────────► │   handler     │
  └──────────────┘                        └──────┬────────┘
                                                 │
                              ┌──────────────────┴──────────────────┐
                              ▼                                     ▼
                     ┌────────────────┐                   ┌──────────────────┐
                     │   embedder     │                   │   Turso (local)  │
                     │  hugot + ORT   │                   │  F32_BLOB(768)   │
                     │  nomic v1.5    │                   │  vector_distance │
                     └────────┬───────┘                   └──────────────────┘
                              │  768-dim                             ▲
                              └──────────────  query vector  ────────┘

The binary itself is CGO-linked (hugot ORT backend + static libtokenizers.a). At runtime the only native surface is libonnxruntime, loaded via dlopen after a SHA256-verified auto-download. Everything else — Go stdlib, tursogo, the model weights — is either statically linked or fetched on first launch against a pinned hash. No system installs. No sudo. If a download drifts from its pinned hash, the run aborts; there is no fallback to an unverified fetch.

Escape hatches for air-gapped boxes:

Default cache paths per platform:

A sibling deadzone.db.release JSON manifest records {tag, sha256, fetched_at}. Startup compares the cached tag against the binary's compiled-in version: match → fire a 3-second freshness probe against deadzone.db.sha256 on the matching GitHub Release, atomic-swap if the remote sha differs (soft-fail to the cache on any network error); differs → fetch the new tag's release and atomic-swap; dev build → fall back to /releases/latest with a server.db_version_dev_fallback WARN. Pre-#197 binaries wrote a single-line tag-only sidecar; the JSON reader still accepts that format and rewrites it to v1 on first probe.

Add a library

Contributor path. End users don't touch this — they just get what ships in deadzone.db.

Not editing YAML yourself? Open an issue via the New issue page and pick Add a library or Refresh a library. The template collects exactly what a registry entry needs.

Editing YAML yourself? Append to libraries_sources.yaml:

libraries:
  # Single-version lib — no `versions` key, urls used as-is.
  - lib_id: /modelcontextprotocol/go-sdk
    kind: github-md
    urls:
      - https://raw.githubusercontent.com/modelcontextprotocol/go-sdk/main/README.md
      - https://raw.githubusercontent.com/modelcontextprotocol/go-sdk/main/docs/quick_start.md

  # Multi-version lib — `versions` expands into one effective lib_id
  # per version (/org/project/1.4, /org/project/1.5, …). {ref} is
  # substituted from each version's ref: field.
  - lib_id: /modelcontextprotocol/go-sdk
    kind: github-md
    versions:
      "1.4": { ref: v1.4.1 }
      "1.5": { ref: v1.5.0 }
    urls:
      - https://raw.githubusercontent.com/modelcontextprotocol/go-sdk/{ref}/README.md
      - https://raw.githubusercontent.com/modelcontextprotocol/go-sdk/{ref}/docs/getting-started.md

Pre-1.0: no Go editing, no recompile. Just edit YAML and re-scrape.

Scrape-via-agent (experimental)

⚠️ The messy half. Works today for non-markdown sources (Terraform providers, mkdocs, GitBook, …), but the LLM→verifier loop is sensitive to input truncation (48 KiB cap), HTML→markdown skill, and verbatim-code matching. Real-world hit rate on dense doc sites ≈ 50%/URL — see #64. Prefer github-md whenever the project ships committed markdown.

Bring your own LLM runtime — Ollama, llama.cpp, vLLM, LocalAI, LM Studio, Groq, OpenAI, anything that speaks POST /v1/chat/completions:

export DEADZONE_AGENT_ENDPOINT=http://localhost:11434/v1
export DEADZONE_AGENT_ENDPOINT_MODEL=qwen2.5:7b
export DEADZONE_AGENT_ENDPOINT_API_KEY=sk-...   # optional

Then add a kind: scrape-via-agent entry to libraries_sources.yaml with a list of page URLs. The downstream pipeline (parse → chunk → embed → store) is identical to github-md; only the markdown source changes.

Guardrails. Every fenced code block in the LLM output is verified verbatim against the source — invented examples drop the doc (scraper.agent_verification_failed), not the run. Missing/unreachable endpoint aborts at startup; no silent fallback.

Local pipeline (contributors)

Two-step bootstrap: toolchain first, then the CGO native dep — kept separate so air-gapped / CI runners with vendored libtokenizers.a can skip step 2 by overriding DEADZONE_TOKENIZERS_LIB.

just bootstrap            # Go 1.26.2 + just toolchain (mise install)
just fetch-tokenizers     # libtokenizers.a — one-shot CGO setup
just build                # CGO + ORT, all packages
just scrape                       # all libs — one artifact folder per lib
just scrape /hashicorp/terraform  # one base lib, every version
just scrape /hashicorp/terraform/1.14   # one exact version
just consolidate                  # merge artifacts/*/artifact.db → deadzone.db
just serve                        # MCP server against deadzone.db

just with no args lists every recipe. Each scrape rewrites artifacts/<slug>/artifact.db + state.yaml in place; consolidate merges all artifact DBs atomically under deadzone.db. Per-lib folders are gitignored; the committed artifacts/manifest.yaml records release history only.

Full registry via CI. gh workflow run scrape-pack.yml -f tag=vX.Y.Z fans out the matrix, consolidates, and uploads deadzone.db to the tagged release. Omit -f tag=… to stop at a consolidated-db cache.

Release flow

Two-phase as of #101 — CI ships binaries, operator ships the DB.

# 1. Regenerate deadzone.db from the committed scraper config.
just scrape && just consolidate

# 2. Tag + push. CI's release.yml builds tarballs + AppImages, creates the release,
#    and auto-bumps the Homebrew tap on release.published.
git tag v0.X.0 && git push --tags

# 3. Ship deadzone.db + deadzone.db.sha256 to the same release.
just dbrelease v0.X.0

# 4. Commit artifacts/manifest.yaml so the release-history trace lands in git.
git add artifacts/manifest.yaml && git commit -m "release v0.X.0" && git push

A stable-tag push fans out fully through CI: release.yml -> chain-release.yml dispatches scrape-pack.yml -> chain-image.yml dispatches docker-publish.yml (one workflow per concern, chained via workflow_run).

Manual Homebrew fallback. The tap auto-bump fires on release.published (#148). If RELEASE_PUBLISH_TOKEN expires and the chain breaks, run it by hand:

gh workflow run update-package-channels.yml -f tag=v0.X.0

Logs

Structured JSON on stderr via log/slog. Stdout is reserved for MCP JSON-RPC on deadzone server.

--verbose on any subcommand adds debug-level detail. On server it logs the raw query (off by default — queries may carry user data). On scrape it adds per-doc scraper.doc_indexed.

MCP client log paths: Claude Code on macOS writes to ~/Library/Logs/Claude/mcp-server-deadzone.log; other clients vary.

Roadmap & contributing

Issues: laradji/deadzone/issues. Scope via milestones. Category via feature / research labels; priority via P1 / P2 / P3.

New library or refresh: use the New issue page and pick the matching form.

Why bother with vectors

Because "how to register a tool" should find the doc that says AddTool, and no FTS5 query will get you there without the human already knowing the answer. Embeddings-first retrieval is the point; everything else is plumbing.

Long-form: docs/research/context7-analysis.md.

License

Apache License, Version 2.0. Third-party attributions in NOTICE.

One important asterisk. Apache 2.0 covers the Deadzone source code, and only that. It does not cover the third-party documentation the scraper indexes — those docs belong to their original authors under their own licenses. Running deadzone scrape is subject to each source's ToS. A pre-built pack is bound by the original content's license, not Apache 2.0. Personal local indexing: fine. Public redistribution: do the homework first.