Give your AI assistant the real docs.

No. The local MCP server your assistant talks to (`helpike-mcp`) makes no network request during a search at all — the single exception in the whole server is the `docs_recommend` tool, which fetches the public pack catalog over HTTPS and caches it for 10 minutes. Query text is written only to `~/.helpike/activity.jsonl` on your own disk so you can audit what your assistant asked, and the Home screen's "Clear history" button deletes it. Search text, file paths and page titles never leave the machine, not even as a hash.

The app sends one GET per day with counters only — packs installed, MB on disk, number of searches, latency percentiles, which MCP clients connected — and the exact payload is published on the telemetry page and frozen by a test so a new field breaks the build. It is declared in section 5 of the terms of use the app shows you at first launch, and it starts only once you accept them; nothing is sent before that, and no search text is ever sent at all. The app itself also talks to the CDN: it fetches the catalog (reloaded every 30 minutes while open) and checks `/download/latest.json` hourly for updates, which CloudFront logs like any download server; the raw log expires after 45 days and the IP is never stored. Setting `DO_NOT_TRACK=1` or `HELPIKE_NO_TELEMETRY=1` stops the daily report.

Not on macOS. The app is signed with a Developer ID certificate and notarized by Apple, and so is the .dmg that carries it, so the disk image opens with no warning and the first launch only asks the ordinary "downloaded from the Internet" confirmation — click Open, once. Nothing has to be stripped by hand: the `xattr -dr com.apple.quarantine` step older instructions asked for is no longer needed. Windows is the one that still complains: the installer has no code-signing certificate yet, so SmartScreen shows "Windows protected your PC" the first time — click "More info → Run anyway". Linux has no equivalent gate; install the .deb or .rpm normally.

Search does, yes: packs are SQLite files on your disk and the query is embedded on your own CPU, so nothing is fetched at query time. Measured on an M3 Pro (docs/BENCHMARKS.md): 7–10 ms to embed a short query, 7.2 ms p50 for a 6,700-vector pack and 64 ms p50 for a 70,000-vector one. You need the network only to download packs, the embedding model and app updates — plus `docs_recommend`, the one tool that reads the online catalog to suggest packs you don't have yet.

No. There is no login, session or cookie anywhere in the product, no API key to paste, and no per-call quota — the search runs on your hardware, so there is nothing to meter. The catalog and the packs are served from a public CDN URL; you can read the whole catalog yourself right now with `curl https://helpike.com/packs/catalog.json`.

The embedding model (BAAI/bge-base-en-v1.5) is a one-time ~416 MB download shared by every pack. Packs are installed individually and most are small: the median pack in the catalog is under 4 MB, Docker is ~10 MB, PostgreSQL ~32 MB, Kubernetes ~57 MB; the largest single pack, Django, is ~480 MB. The full catalog is 1,389 packs and 10.9 GB, but you only download the ones you pick, and any pack can be deleted from the app.

Any MCP client that speaks stdio. The app hands you a copy-paste prompt that registers the server at project scope, with the exact form for Claude Code (`claude mcp add --scope project`), Cursor (`.cursor/mcp.json`), VS Code (`.vscode/mcp.json`) and Antigravity (`.agents/mcp_config.json`); any other client uses its equivalent project-scoped config. Your client launches `helpike-mcp` itself, so the desktop app does not need to be open for searches to work. Four tools show up: `docs_search`, `docs_page`, `docs_packs` and `docs_recommend`.

Context7 is a hosted service metered per call — when we checked its pricing page (July 2026) the free tier was 1,000 calls/month and Pro was $10 per seat/month for 5,000 — and even its self-hosted enterprise build still reads Upstash's public catalog, so it is not air-gapped (issue upstash/context7#320 asking for that was closed without implementation). Helpike ships the index to you instead: packs are files on your disk, queries never leave the machine, and there is no per-call cost or cap. The trade-off is honest — you download the ~416 MB model plus the packs you choose, instead of getting any library instantly from a server.

docs-mcp-server (MIT) is a solid tool and uses essentially the same engine — SQLite + sqlite-vec + FTS5 hybrid search — but it expects you to crawl each site and bring your own embedding provider. Helpike's packs arrive already crawled, chunked and vectorized with bge-base-en-v1.5, each verified by the sha256 published in the catalog, so adding the Kubernetes docs is a download instead of a crawl-plus-embedding job. If you enjoy running your own indexing pipeline, that project is the honest recommendation.

Sources are re-crawled on a per-source schedule — daily for fast-moving docs, weekly to monthly for stable ones — and only pages whose sha256 changed are re-embedded, so rebuilds are cheap and frequent. The app reloads the catalog every 30 minutes while open and marks any installed pack whose published sha256 no longer matches yours as "Update"; refreshing it is one click. Every chunk stores the source URL and its availability/version, so the assistant can cite the page it actually read.

The current beta (2.5.0) builds are macOS on Apple Silicon (.dmg, aarch64), Linux x86_64 (.deb, .AppImage, .rpm — Ubuntu 22.04+, Debian 12+, Fedora) and Windows 10/11 x64 (.exe, .msi). There is no Intel Mac build and no Linux arm64 build today. On Windows, a missing-DLL error at startup means you need the Visual C++ Redistributable (https://aka.ms/vs/17/release/vc_redist.x64.exe); on Linux prefer the .deb/.rpm, since the AppImage needs FUSE 2 and changes the server path on every run.

The catalog currently holds 71 documentation sets and 1,389 packs, built only from official documentation sources — Docker, Kubernetes, React, Go, Python, PostgreSQL, Django, Rust, PHP, MDN, Symfony, Terraform/OpenTofu and more; huge sets are split so each pack stays fast (Apple is 405 per-framework packs, AWS is 299 per-guide packs). You can check before installing anything: `curl https://helpike.com/packs/catalog.json`. Once connected, `docs_recommend` reads your project's dependency files (package.json, go.mod, requirements.txt, build.gradle…) and tells you which packs match and which you already have.

Write to support@helpike.com — the same address handles support, bug reports, security disclosures and telemetry deletion requests, and it is published in machine-readable form at /.well-known/security.txt as RFC 9116 expects. For a security issue, include reproduction steps and estimated impact; the commitment is a first response within 72 hours, and please do not open a public report before a fix exists. To have telemetry rows deleted, send the install id the app shows on its usage-data screen. There is no account or ticket system to sign up for.