Typed pipelines for coding agents

One Go binary and your repo. A workflow is a YAML file; each step is a prompt, scoped skills and tools, and a JSON-schema output contract, and the engine won't advance until the output validates. Build the workflow inside the agent you already use, then run it against any backend.

Status: v0.1.0 release in progress; install from source today.

Docs at wfnexus.deemwar.com → · Get the alpha

How it works

A step is a contract

Each step is a prompt, a scoped set of agent skills and tools, and a JSON-schema output contract. The engine refuses to advance until the step's output validates, and a failed validation comes back to the agent as a tool result so it corrects itself mid-turn.

Free dry run

wfx dryrun checks whether a workflow would run here with no model call, no clone and no write.

Cost per step, while it runs

Tokens and cost are aggregated per step and written while the step runs, not at the end.

Human approval gates

A workflow stops for a person where it says so: approve, reject, give input, retry or cancel.

Live view

The agent's tool calls stream to the UI over SSE, and every step's state is stored (SQLite or Postgres).

Runs on the agent you already use

The same workflow runs on claude, codex, copilot, opencode or devin, or on your own HTTP model.

Self-host it

One Go binary and your repo, on SQLite and a folder with no configuration. Or Docker Compose: wfx-server plus Postgres.

Git is the registry

wfx publish <file.yaml> --version <semver> --to <git remote> writes the workflow, its skills and MCP declarations into a git repo, tagged and immutable.

Alpha

Alpha. One-time licence, price on request.

The self-host build stays open source (0.x, MIT). The team server is the alpha. Need a workflow built for you? That is a service at EUR 60/hour, negotiable: see consulting.

Get the alpha

Or email us.