deadzone

Overview

Deadzone is an offline semantic search engine for library documentation. Users can query in natural language (for example, "how do I register a tool?"), and it uses vector embeddings to find relevant passages in the documentation (for example, where AddTool is mentioned), which traditional grep- or FTS5-based indexing cannot do.

It is officially described as "Context7 but with the internet turned off": a Go binary that communicates with AI clients over stdio using the MCP protocol, with all data remaining on the user's laptop.

Main capabilities

• Natural language first: Uses the nomic-embed-text-v1.5 embedding model and cosine distance for retrieval, not FTS5.
• Local index: The index is stored in a local Turso file, with no account, no API key, and no outbound network requests on the hot path.
• Single binary: All functionality is provided through the deadzone subcommand, with no need for pip, npm, or docker compose.
• Version locking: The database is bound to the binary version. After upgrading the binary, the database is fetched again on the next startup; if not upgraded, the cache continues to be used.
• Index fetch on first startup: On first startup, it fetches the deadzone.dbmatching the current binary version, performs SHA256 verification, and caches it in the platform data directory. After that, startup is offline.
• MCP over stdio: Can be integrated as an MCP server with MCP-capable clients.
• Zero telemetry: No telemetry data.
• Container support: Provides multi-architecture container images, with the database baked into the image, no volume mount required, and offline usability immediately on startup.
• Air-gapped environment support: Provides environment variables and default cache paths to facilitate offline deployment.

Usage

Installation

Binaries, tarballs, or AppImages can be obtained from the releases page, and installation is also possible via Homebrew. Container images are provided at ghcr.io. All methods ultimately produce the same deadzone executable.

MCP client configuration

Native binary method:

./deadzone server

Container method: The image already includes deadzone.db, with no volume mount required.

Contributor path

You can add a library source by editing libraries_sources.yaml , then run the crawl. You can also create a new issue on GitHub and select the "Add a library" or "Refresh a library" template.

Experimental crawl proxy

Crawling is supported through an LLM runtime compatible with POST /v1/chat/completions (such as Ollama, llama.cpp, vLLM, LocalAI, LM Studio, Groq, OpenAI), but the official documentation notes that this process is sensitive to input truncation, HTML→markdown conversion, and verbatim code matching, with an actual hit rate of about 50%/URL. It is recommended to prefer github-md.

Local pipeline

Two-step bootstrap: first install the toolchain, then handle CGO native dependencies. Manage with the just command. Each crawl rewrites artifacts/<slug>/artifact.db and state.yaml，consolidate atomically merges all databases into deadzone.db.

Applicable scenarios and limitations

Applicable scenarios:

• Semantic search of library documentation in local, offline environments.
• Need to integrate documentation retrieval into AI clients (via MCP).
• Scenarios with high data privacy requirements and no desire to send queries to the cloud.

Limitations:

• Does not override the copyright of third-party documentation; prebuilt packages are subject to the license of the original content.
• The crawl proxy is an experimental feature with a limited hit rate.
• Windows is blocked upstream (missing libtokenizers.a), so WSL must be used.
• AppImage requires FUSE v2; some minimal servers require additional installation or the unpack parameter.
• The category identifier is ai-office-tools, but the actual use is more toward a documentation retrieval tool.