An agentic coding workspace that runs on your models.
Describe a feature or point it at an existing repo. A pipeline of specialised agents researches, plans, writes, validates and repairs the code — against a real compile, test and lint loop, not a guess.
It runs on your own machine, against your own model server. Point it at a local Ollama or vLLM, or at an Inferix inference endpoint — the code never leaves your boundary for a third-party AI provider.
Currently in private early access. Request an invite and we'll be in touch.
researcher · indexed 412 files · 3 project rules loaded
orchestrator · 7 tasks queued — backend first
coder · wrote api/routes/reports.ts
coder · wrote app/reports/page.tsx
auditor · build ok · lint ok · 2 tests failing
debugger · root cause: missing await in loader
coder · applied fix · re-running audit
auditor · all green — handing back
Not one prompt — a pipeline
Each turn runs through a state machine of dedicated roles. The work only comes back to you once validation is actually green.
Researcher
Reads the workspace tree, runs semantic code search over your repo, and picks up project rules before anything is written.
Enhancer & orchestrator
Turns a rough request into a concrete spec, then breaks it into an ordered, file-by-file task queue you can edit before it runs.
Backend & frontend coders
Write the actual files, with a backend-before-frontend gate so a UI is never generated against an API that does not exist yet.
Auditor
Runs the real thing: backend compile and tests, frontend build, lint, a headless-browser smoke pass, a security review and an accessibility check.
Debugger
On a failure it reads the actual error output — and, with a vision model, the rendered page — then hands the coder one targeted fix instead of raw stderr.
A workspace, not a chat box
Sandboxed terminals
Commands run in a per-workspace container with three autonomy tiers: confirm everything, confirm risky commands only, or run unattended.
Real git workflow
Branch, stage, commit, push and open a pull request from the workspace. Commit messages are drafted from the actual diff.
Per-role models
Eleven agent roles, each with its own model, temperature and endpoint. Run the coder locally and the debugger on a bigger remote model in the same session.
Your code stays put
It runs on your machine against your model server. No third-party AI provider sits in the path, and nothing is uploaded to be indexed.
Bring your own compute
An agent run is a lot of inference — research, two coders, an auditor and a debugger, plus a vision model when a UI fails. Coder speaks the OpenAI-compatible API, so every role can point wherever you want it to.
Per-role routing
Each role carries its own model, endpoint and key, so a heavy debugger can run remotely while everything else stays on your machine.
Try it on your own codebase
Inferix Coder is in private early access while we harden it for other people's machines. Tell us what you're building and we'll get you an invite.
Request early access