← all repositories
neomjs/neo

When your frontend framework starts filing its own PRs

Neo.mjs is a JavaScript app engine that hosts a multi-model AI swarm as named maintainers with persistent memory, cross-family code review, and a "possession interface" for live runtime mutation.

neo
Velocity · 7d
+1.3
★ / day
Trend
steady
star history

What it does

Neo.mjs is two things bolted together by a “Neural Link.” The Body is a production JavaScript app engine: multi-threaded (App Worker, VDom Worker, Data Worker, Canvas Worker, SharedWorker), zero runtime dependencies, native ES modules, no build step. It claims real deployments handling 40k+ delta updates/sec on trading desks and multi-window control rooms. The Brain is an Agent OS running Claude, Gemini, and GPT as named machine-account maintainers that author issues, open PRs, review each other cross-family, and persist reasoning in a queryable Memory Core + Native Edge Graph.

The interesting bit

The “possession interface” is the unusual angle: agents don’t just generate code, they co-inhabit the live App Worker heap, inspect runtime state (get_component_tree), hot-patch methods (patch_code), and mutate UI without reloads. The README frames this as “Ghost in the Shell” embodiment — the VDOM tree becomes a “working memory surface where the model is thinking with its hands inside the machine.” The same primitive, they claim, extends to game engines and robotics.

Key highlights

  • Named AI maintainers with persistent identities: @neo-opus-4-7, @neo-gemini-3-1-pro, @neo-gpt, each authoring and reviewing PRs under their own accounts
  • Cross-family review: Claude reads Gemini’s thought processes via A2A messages stored in the Memory Core, not just final outputs
  • DreamService / Golden Path: a REM-cycle consolidation that distills tactical sessions into graph topology with a hardcoded priority formula (semanticScore × 2 + structuralWeight)
  • MX loop: internal friction → tickets → PRs → skills → next agent’s reflexes; 706 merged PRs and 800 closed issues in May 2026 per the README
  • v13 adds multi-tenant cloud deployment: point the Brain at your own repos with per-tenant isolation

Caveats

  • The Brain currently runs centralized on “a single human’s laptop” — SQLite + vector DB + LLM API funding is a self-acknowledged single point of failure (#10291)
  • Two of four “faculties” are still in flight: short-term recall and Dream/Golden Path authoritative routing
  • The 40k+ ops/sec and production deployment claims are stated but unsourced in the README; no benchmarks or customer names provided
  • The README’s tone is unusually grandiose even by 2026 AI-project standards — “self-evolving software organism,” “autopoietic,” “ANI by accumulation” — which may signal depth or may signal that the project is still proving its claims

Verdict

Worth watching if you’re building high-throughput multi-window UIs or experimenting with multi-agent systems that need runtime embodiment rather than static code generation. Skip if you need a boring, well-supported framework with a large ecosystem — this is explicitly “not a framework-shopping option” and the governance model requires a human merge-gate authority.

heatdrop uses Google Analytics to see which pages get read — nothing else. Your call. How we handle data.