FAQ

No. Analysis runs locally against your files and your language servers. No source code or analysis content is uploaded. License verification runs at most daily against a 3-day signed offline lease. This is the complete license request:

POST /v1/license/verify

{
  "api_key": "ce_xxx",
  "machine_id": "persisted-random-device-uuid",
  "cli_version": "0.1.0",
  "os": "darwin",
  "os_version": "14.5",
  "telemetry": {
    "tool_calls": {},
    "tokens_saved_estimate": 0,
    "languages_used": []
  }
}

Two ways to verify that yourself: (1) cut the network after one successful verification — analysis keeps working for the full lease window, offline, and then tells you exactly why it stopped; (2) watch the wire with lsof -i or Little Snitch during a session, then inspect each destination and payload.

Because the codebase is the product being commercialized — that is the honest answer. The trust questions that actually matter (what leaves my machine, what happens if you disappear) are answered with verifiable mechanisms instead of source access: the wire capture above, the offline-lease test you can run yourself, and a published discontinuation commitment (below).

License enforcement and abuse control for a free preview. Key issuance is instant — GitHub sign-in, about 30 seconds — and the preview is free.

Exactly the fields in the request above: per-tool call counts, a bytes-based token-savings estimate, and language IDs. Never code, file paths, repository names, or symbols.

If the Community Preview service is ever discontinued entirely without a successor service — any maintained offering of this technology by us or by any acquirer or successor operator, regardless of branding or pricing — we commit to shipping a final build that runs without license verification. Separately: if paid tiers arrive, preview users get 30 days' notice, and the binary keeps working offline for the full lease window during any outage.

Context Engine combines a polyglot structural layer with live language-server semantics. It turns files into addressable modules, APIs, implementations, types, dependencies, and usages, then lets an agent fetch only the view it needs. A filesystem watcher routes changes to the relevant language servers so future requests reflect the synchronized workspace rather than a pre-built semantic snapshot.

The MCP server is a thin transport adapter over an embeddable Rust core. The current build ships 324 bundled Tree-sitter grammars (macOS-verified; Windows verification pending) and 147 default LSP configurations covering 143 language IDs, backed by ~5K tests and ~1,300 doctests. That infrastructure is what makes one small agent-facing interface work across languages, workspaces, and exact dependency versions.

No public release archives are available yet. After the first verified immutable release, exact-version archive and checksum links will appear on the download page.

The Community Preview is free. We will announce any change to availability or terms in advance. There is no automatic conversion to a paid subscription.