Local Multi-Agent LLM Platform
A modular, offline-capable multi-agent platform (OpenAI-compatible gateway, control plane, autonomy engine, RAG, and IDE integration) with an automated adversarial peer-review step and zero hosted-LLM dependencies.
- ~32k
- lines of code estimated
- 0
- hosted-LLM dependencies measured runs fully local / offline
Private infrastructure, not published.
Context
I wanted a senior/junior agent team I could run entirely on my own hardware: no cloud, no per-token bill, no data leaving the building. It had to actually maintain and extend itself under review.
What I built
I built an OpenAI-compatible gateway fronting two locally-served models, with a control plane and a browser dashboard for full visibility. Automated adversarial peer-review: a generator reviews recent commits before merge, and larger changes get a separate multi-dimension review pass across correctness, security, performance, and API contract, so generated work is challenged before it is trusted. An autonomy engine gates every agent action with an allow/approve/block policy, backed by least-privilege fail-closed tool bridges and per-project privacy isolation. A deterministic competence benchmark (keyword-rubric, no LLM judge) measures and gates quality instead of asserting it.
What holds up in practice
It keeps working, catches its own mistakes before merge, and never depends on a hosted model to run. The second model lives on the RTX 3090 node, which is down with a hardware fault until I repair it, so the platform runs on one model for now.
The platform itself is private infrastructure and is not published; the RAG and retrieval components extracted from it are the public repositories linked elsewhere on this site.