c8c turns your AI skills into flows with quality gates and human checkpoints. Works with Claude Code, Codex, and OpenClaw.
No account needed. Works with your existing subscription.
You already do this by hand
You run plan → code → review → test → ship with Claude Code every day.
The work is good. The handoffs still happen one manual step at a time.
Every handoff means another prompt, another paste, another place where context and artifacts disappear into scrollback.
The system only moves forward when you manually pass output to the next step, wait for it, and decide what to do next.
What c8c does
Run a multi-step flow, inspect each stage, recover from the one that failed.
01
Pick a template or describe what you want. c8c builds the flow and runs it. You come back to results, not to the next manual step.
Start from a template or generate the flow from a task description.
Stages execute in order with fresh context.
You come back to named results, not another handoff.
02
Evaluator gates score output at every step. Below threshold → auto-retry from the step that failed. Stage fails entirely? Rerun from that stage, not from scratch.
03
Approval gates pause before risky stages. You can also edit output before continuing, fill in a human-task form, or resume a paused flow hours later from the inbox.
Approval is only one checkpoint type. Editable and structured human tasks are also first-class.
Paused runs keep their state instead of forcing another reconstruction pass.
The inbox shows what's blocked so you can unblock it on your schedule.
04
Flows are saved as YAML in your repo. Batch one process across 50 inputs with a multi-run dashboard, then share the same YAML for someone else to run in 30 seconds.
YAML definitions in your repo — readable, versionable, diffable.
Batch the same flow across many inputs with a multi-run dashboard.
Hand the YAML to a teammate. They run it in 30 seconds — no setup.
How it works
c8c handles sequencing, evaluator gates, and pause points in between.
Choose Dev Process template — or describe the job in one sentence and let the agent generate the flow for you.
Start from a readable template instead of a blank graph.
Adapt the flow later — it's a YAML file you can read and edit.
draft
Stages run in sequence. Quality checks happen between steps automatically. The flow pauses only at approval gates and human-task forms you placed.
A failed stage reruns from its checkpoint — prior stages keep their results.
Resume paused work later without losing the run state.
run
Named outputs at each stage: spec, plan, review, PR. Per-stage logs, token usage, duration, and cost stay inspectable after the run ends.
Not a blob — structured outputs with full observability.
You inspect evidence and outputs at each stage.
results
In practice
Task: 'Add photo upload for sellers.' 11 files changed. 9 tests added. 3 bugs caught before deploy.
Works with what you have
No account. No migration. No need to replace the skills or provider you already use.
No. Built-in templates work with raw Claude Code prompts. Bring your own skills later.
No. c8c doesn't replace your skills. Your .claude/commands/ stay untouched. c8c connects them.
Claude Code, Codex, and OpenClaw. Use whichever you're subscribed to.
Pick a template, paste your task, press Run. First artifact in 8 minutes.
Open source, MIT license. Runs locally on your machine. No account, no server, no fees.
Your workflows are YAML files in your repo. Portable by design.
Yes. `c8c-workflow run`, `resume`, `rerun-from`, `hil approve` — the same flow model works headless. Pipe it into CI, cron, or OpenClaw.
CLI
Portable by design
Same YAML runs in the desktop app, in CI, or headless. No separate format per runtime.
You, if...
Plan, code, review, test, ship — and the orchestration overhead slows you down.
Run the full loop once. Come back to the spec, plan, review, and PR — not to another copy-paste.
It sequences your AI work until a step fails, someone else needs context, or you have to rerun one part cleanly.
c8c keeps the same process readable enough to operate without hand-built glue.
You already know the harness pattern and still run the quality gates by hand between steps.
Your skills stay the same. c8c adds evaluator gates between them and a desktop UI to review results.
Why this matters now
Your skills already work. The missing piece is what runs between them — quality checks, sequencing, approval gates.
The gap is between them — where context drops, output goes unchecked, and failures compound before anyone notices.
Add one evaluator-gated retry per step. Per-step rate: 85% → 98%. End-to-end: 44% → 89%. The model didn't change. The harness did.
Harness workflows today run headless — shell scripts, CI pipelines, tmux grids. c8c lets you see results, approve stages, and inspect artifacts without switching back to the terminal.
'Harness' coined by Mitchell Hashimoto (Feb 2026). Adopted by OpenAI, Anthropic, Stripe, Martin Fowler. c8c adds the review surface to that pattern.
Social proof
c8c connects them into flows you can run, review, and share.
superpowers, GSD, impeccable, and gstack show real demand for harness-style workflows. c8c focuses that demand on runtime control, review, and human checkpoints.
Install
Open source. MIT license. Free with your existing AI agent subscription.
Star on GitHub to follow development, share workflows, and contribute templates.