productionflagship
Brainiak
Local-first memory system and MCP tool server for coding agents — with certified on-device AI.
Brainiak gives any MCP-compatible coding agent durable memory, hybrid semantic search and a safe patch pipeline — running entirely on your machine. v0.3.0 ships 25 production tools, a certified embedding model (100% Recall@1 on GPU/CPU/NPU) and a model control plane with SHA-256-pinned installs and consent gates. No cloud, no data egress.
●
Solution
Brainiak sits between IDE agents and local repositories. It indexes workspaces, records episodic and decision memory, enforces a LocalOnly data-flow policy, and delivers context packs on demand — while its certified embedding model (Qwen3-0.6B) runs on-device via llama.cpp, OpenVINO or CPU at 17–22ms per query. Every model install is SHA-256-pinned, consent-gated and audited; every edit flows through preview_patch → run_health_checks → apply_patch. Details, screenshots and install guides at brainiak-landing.vercel.app.
Highlights
- ▸25 production MCP tools over JSON-RPC (stdio & HTTP), schemas served live by the Rust registry.
- ▸Hybrid retrieval (BM25 + on-device embeddings) certified at 100% Recall@1 — GPU 17ms, CPU 22ms, NPU 91ms.
- ▸Model control plane: SHA-256-pinned installs, consent gates, idempotent lifecycle and per-device benchmark evidence.
- ▸Safe patch pipeline (preview → health checks → apply) with compiler-output parsing.
- ▸LocalOnly data-flow policy, fail-closed runtime, workspace isolation and a full audit trail.
- ▸~70% fewer agent tokens per retrieval task.
Stack
RustTokioAxumSQLiteMCPReact 19TypeScriptllama.cppOpenVINO