Nyx
A self-hostable, autonomous multi-agent system that runs a team of AI agents to build software, research, and automate, on your own models, on a spine engineered so it cannot get stuck.
Most AI tools wait for you to type, then do one thing and stop. Nyx is different: it is an autonomous orchestrator that runs a whole team of specialist agents on its own, around the clock, on a machine you control. You give it a goal, it decomposes the work, fans it out across agents, checks the results, and keeps going until it is done, recovering itself when something fails instead of silently hanging. It is the part you would normally hire people for: a system that takes a fuzzy objective and turns it into finished, verified work without supervision.
Think of it as a small studio that never sleeps. An Assistant takes the request and delegates. A Coder builds. A Researcher digs. A Reviewer opens a real browser and checks the result looks right and works on every screen. The whole studio runs on a self-healing spine, on the models you choose, and you can point more of your own devices at it as workers.
One system, several jobs. All of it self-hosted and runnable unattended.
build software
Give it a goal; it decomposes into increments, fans out workers, verifies, and commits each one, overnight, no babysitting.
do anything
Chat, research, automate, code. It understands the request, remembers you, and routes the work to the right agent.
never stuck
A fail-open controller that, by design, cannot wedge: a stuck task is skipped or retried and the run always finishes.
your stack
Runs on your machines on a configurable model Route (DeepSeek, Groq, and others), so it is private and cheap.
a teammate
The Coder watches a repo's issues, fixes them through the verify loop, and opens a PR that closes the issue.
scales out
One brain, many hands: point more of your own devices at the coordinator and they pick up work as it comes.
Two orchestrators over a set of doers. The conductors coordinate and aggregate; the doers produce the work. See inside the agents.
orchestrator
The front door. Understands, remembers, routes, and aggregates. Does no grunt work itself.
doer
Reads, patches, runs, and tests in a sandbox until the change is verified. Also a GitHub teammate.
doer
Fans out searches, reads and ranks sources, returns a synthesized, cited answer.
doer
Opens a browser, clicks through, checks desktop and mobile, and sends fixes back to the coder.
One coordinator owns the work and cannot wedge. Agents run on a swappable harness. Workers can live on any device.
Everything swappable is config, not a rewrite. This is how Nyx stays cheap, fast, and not locked to any vendor.
The agent runtime. Switch with NYX_HARNESS. Migrate gradually, fall back instantly.
Planner, worker, vision, image, embedding each pick their own model. DeepSeek to plan, a fast model for the rest.
The assistant routes and aggregates; the coder, researcher, and reviewer do the work.