Axiomux

Inspiration

Coding agents increasingly offer many models, but switching providers often means switching the entire operational context: credentials, history, memory, tools, permissions, and recovery semantics. Axiomux separates the model-execution substrate from the agent's continuity and authority.

What it does

Axiomux is an MIT-licensed control plane for coding agents. It unifies provider login and model selection while keeping projects, steering, governed memory, quarantined skills, effect approvals, fallback, and tool-less multi-model Fusion under one operator-controlled runtime.

OMP runs as a pinned, digest-verified external process for supported authentication and model execution. Axiomux is not an OMP fork or extension: it owns the project registry, sessions, steering, follow-up queue, local tools, effect journal, memory, reusable skills, workflows, fallback, Fusion/MoA, and bilingual local UI. Neuralwatt is implemented as a native provider and appears in the same catalogue as other flows.

Memory is stored locally with provenance, review state, feedback, relations, and bounded retrieval. Automatic learning cannot silently activate code: it creates a candidate, requires human review, converts it into quarantine, requires a second approval, and records evaluation. Consequential tool effects are journaled before execution. Cross-provider fallback stops once an effect makes replay unsafe. Fusion consultants are tool-less and failures are isolated; only the primary agent can request actions.

How we built it

Axiomux uses TypeScript, Node.js 24, SQLite/FTS5, and a bilingual local HTML/CSS/JavaScript UI. The judge package includes a deterministic offline demo that executes real production components without provider credentials, network access, or user-state mutation. It hands one local session from an exact Provider A route to Provider B while retaining the prior answer and approved project memory. It then proves that a pre-effect provider outage can fall back, while a post-effect outage stops without replaying the completed action.

ChatGPT accelerated repository analysis, architecture, implementation, security hardening, regression tests, UI work, release repair, and evidence packaging. I made the core product and safety decisions: create a new harness rather than a mod; keep strict isolation and explicit approval; preserve fallback without replaying effects; make Fusion arbitrary but tool-less; and separate general engineering improvements from submission-only work. A separately recorded live check returned the requested exact token through openai-codex/gpt-5.6-sol.

Challenges

• Reusing OMP's login surface without becoming an OMP fork or inheriting extension authority
• Keeping fallback useful while preventing retries from repeating external effects
• Making memory and self-learning valuable without treating imported text or generated code as trusted
• Packaging the compiled CLI and installer so a fresh consumer installation needs no TypeScript toolchain
• Distinguishing catalogue capability, successful authentication, model availability, and a real turn in every public claim

Accomplishments

• One continuity and governance layer across heterogeneous model providers
• Deterministic, credential-free continuity handoff and effect-aware fallback using shipped production components
• Two-stage human review for learned skills
• Tool-less Fusion with partial-failure isolation
• Native Neuralwatt integration without enabling third-party OMP extensions
• A self-contained npm tarball plus explicit licensing, provenance, threat model, and known issues

What we learned

Continuity and authority should belong to the operator's local harness, not to whichever model happens to answer the next turn. Useful fallback must be effect-aware, and automatic learning is only safe when candidates remain reviewable and quarantined.

What's next

Finish a dated provider smoke-test matrix; add formal SQLite migrations, backup/restore tests, and optional encryption at rest; validate Windows and Linux packages in remote CI; strengthen isolation for browser/desktop automation; and benchmark recall, Fusion quality, latency, cost, and crash/replay behavior.

Axiomux is an honest alpha, not a claim of universal provider support or a production sandbox. Catalogue entries, account entitlement, and successful model turns are reported separately, and browser/desktop adapters remain experimental.

Built With

• chatgpt
• node.js