The questions security reviews ask.

Open core. The engine, the gan CLI, and the MCP server form the open half — free forever, fully functional offline, fully functional with AI disabled. Ganforge (the hosted index, hammer compute offload, and pull-mode sync) is the commercial half. The exact OSI-approved license for the open half will be named with the public source release; the boundary between the halves is already fixed and documented, and it will not creep.

Local: nothing. Embeddings are computed on-device, search runs locally, and the whole system operates with the network cable unplugged. Team: repository content sent for indexing and your query text reach the Ganforge index service, scoped to your tenant — that is what "hosted index" means, and we would rather say it plainly. Enterprise pull mode: vectors are computed once for the org and sync down to each machine; queries then run locally, so query text never leaves your infrastructure.

The bridge round-trips byte-perfectly in both directions — gan import then gan export reproduces the original repository exactly, and that property is tested, not aspirational. The CLI is a 1:1 drop-in: the commands your hands already type. Repos can run dual-mode with .git and .ganvil side by side, so your canonical home can stay GitHub while agents query gan mcp locally.

Everything still works. Structure tools (outline, callers, impact, symbol_diff and the rest) are computed from parsing and graph analysis, not from a model. Search degrades to lexical ranking — the same code path, one rung down the ladder. AI is additive in Ganvil, never load-bearing; that is an architectural invariant (ADR-0005), not a feature toggle.

No. The MCP surface is read-only with one exception: propose, which is off by default and, when enabled, files a proposal into a review queue — it cannot touch canonical history. Hosts can additionally require that proposals carry an Ed25519 signature from one of the repository’s trust anchors. AI can suggest; only you can make it true.

No. The compute ladder tries the neural engine first (Apple ANE via CoreML), falls back to the GPU, and can be pinned to CPU where that makes sense. If you would rather spend none of your own cycles, the hosted hammer tier takes the indexing work. And the zero-AI mode needs no accelerator at all.

You always know. Every tool response carries a freshness stamp: when it was captured, which head it was read against, and the guarantee that it is exact for that state — good until the next mutation. Tools never block waiting for an index to catch up; they answer from the best available index and tell you exactly what that answer is worth.

Sidecar tools proved that agents with a code graph beat agents without one. But a sidecar is a second source of truth: stale between runs, blind to history, gone when the folder moves. Ganvil computes the graph inside the VCS — updated sub-second per file on every change, keyed by OID so it is queryable at any commit, not just the last time an indexer happened to run.