wfnexus is to coding agents what n8n is to SaaS APIs, but it is aimed at developers. Its three building blocks are agent skills, graph steps and state and memory.
A workflow is one YAML file. Each step is one of two things:
run: shell node.The engine runs each step as a toolnexus agent. It does not move on until the step’s output validates against its schema. It saves every step’s state, and it stops for a human wherever the workflow says to.
Each run tells you what it found: steps write a summary and an outcome, and the run page leads
with the verdict. Workflows can hold one active run per group (concurrency:), and a workflow that
fails to load is skipped by name instead of taking the server down.
It runs anywhere one binary runs. wfx-server boots on SQLite with no config, listens on loopback,
and you reach it from your laptop through an SSH tunnel (wfx context add … --via ssh://…). Model
keys live in its encrypted env store, and a provider that is down falls back to the next one. See
Self-host on one Linux box.
It is not a bug-fix product. The repository started as a bug-fixer, and templates/bug-fixer.yaml
is still one template among eleven. ADR 0010 in the roadmap
records the decision: a workflow platform, not a bug fixer.
wfx templates # starting points: dead-code removal, spec drift, prod watch…wfx new test-gap-finder --as my-tests # copy one into a workflow of your ownwfx dryrun my-tests # would it run here? no model, no repo, no writeswfx run my-tests -i repo_path=. -f # run it and follow the logwfx approve <run-id> # the human gateSteps are agents with contracts
A step is a prompt, a scoped set of skills and tools, and an output_schema. Validation
runs inside the agent loop: submit_output is a tool, so a validation failure goes back
to the model as a tool result and the model fixes it in the same turn.
Any backend
The same workflow can run on an HTTP model (OpenRouter, ollama and others), a CLI agent (claude, codex,
opencode) or an ACP agent (devin, opencode, claude, codex). The shipped registry has 32 providers,
and a step whose provider cannot be reached falls back to the next one in its chain.
Calibrated classifiers
decide: asks typed questions of a small calibrated classifier (TypeSafe jev) before an
agent runs. Its verdicts come in three bands: no, uncertain and yes. An
uncertain verdict goes to a person.
Git is the record
When a project lives in git, every save or delete of a workflow becomes a branch and a pull request. A classifier reviews it, and a person approves or rejects it.
A real Claude Code session, unedited except for speed. Claude builds a five-step workflow, runs it on a free model, stops at a human approval gate, and publishes it. See the demo page for chapters and for what went wrong along the way.
wfnexus · one real Claude Code session · no voice, subtitles on