The problem
You run plan → code → review → test → ship with Claude Code every day. Each step is a separate skill, a separate session, a separate copy-paste. You are the orchestrator — the slow loop between steps that otherwise work fine on their own.
Each step passes at ~85%. Chain 5 steps: 44% end-to-end. Without quality checks, errors compound silently. A batch job wrote 847 bad rows to production — zero error signals.
You might already use gstack, superpowers, or GSD — 135K+ GitHub stars combined prove the pattern works. But they run in the terminal, have no quality checks between steps, and die when you walk away.
What c8c does
c8c turns AI skills into flows you can read, run, and control.
- Run until it needs you. Skills execute in sequence. Evaluator nodes catch failures and auto-retry from the step that failed. You intervene only at approval points.
- Rerun from state, not from scratch. A step fails? Resume from that step. The rest of the flow stays intact. Come back hours later — the state is durable.
- Human loop beyond approve/reject. Approvals, editable review points, human-task forms, inbox with timeout policies. You decide on your schedule.
- 50+ built-in flows. Dev flow, code audit, content pipeline, competitor analysis, cold outreach, UI polish — pick a flow from the library, paste your input, run it.
- Full observability after the run ends. Per-node logs, token usage, duration, active step, typed results — inspectable at any point, not just during execution.
Evaluator checks
Score output against criteria. Below threshold → auto-retry from any upstream step. Same model, better harness: 42% → 78% end-to-end success rate.
Parallel branches
Split work into parallel paths. Merge with configurable strategies: concatenate, summarize, or select best. Fan out 20 competitors, merge into one brief.
Batch processing
Run one flow on 50 inputs. Multi-run dashboard tracks each. Failed items retry individually. Export results as CSV or JSON.
YAML in git
Flows are portable YAML files. Commit them with your code. A teammate clones the repo and runs the same flow. No config, no account.
CLI runner
c8c-workflow run, resume, rerun-from, hil approve — same flow model, headless. Pipe into CI, cron, or OpenClaw for Telegram-triggered runs.
Desktop-first privacy
Everything runs on your machine. No cloud accounts, no data leaving your laptop. Free with your existing Claude Code, Codex, or OpenClaw subscription.
You, if...
- You run Claude Code skills by hand every day — plan, code, review, test, ship — and the orchestration overhead slows you down.
- You've built a bash script or tmux grid to sequence your AI work, and it keeps breaking.
- You use gstack, superpowers, or GSD and run quality checks by hand between steps.
Quickstart
Download the latest .dmg from Releases, or build from source:
git clone https://github.com/bluzir/c8c.git
cd c8c
npm install
npm run dev
Requirements: Node.js 20+, and at least one of: Claude Code CLI, Codex CLI, or OpenClaw.
macOS note: The app is not code-signed yet. On first launch:
xattr -cr /Applications/c8c.app
Or right-click → Open → Open.
No custom skills needed to start. Built-in library flows work out of the box. Bring your own skills later.
How it works
Input → [Skill] → [Skill] → [Evaluator] → pass → [Approval] → [Output]
↓
fail
↓
[Retry from step N]
8 node types cover every pattern:
| Node | What it does |
|---|
| Input | Entry point — text, URL, directory, or batch data |
| Skill | Runs a provider-backed skill with a specific model and prompt |
| Evaluator | Scores output against criteria, auto-retries from any upstream step on failure |
| Splitter | Fans out into parallel branches |
| Merger | Combines parallel results back into one |
| Approval | Human approval — review, edit, approve or return before continuing |
| Human | General human-task form — collect input, decisions, or structured data mid-flow |
| Output | Final result with named results |
FAQ
How is c8c different from Claude Code or Codex?
c8c uses those tools. It chains their skills into flows with quality checks, approvals, and per-step observability. Claude Code does the work; c8c runs the flow.
How is c8c different from n8n?
Complementary, not competitive. n8n handles triggers and integrations across 1000+ services. c8c handles the AI quality layer: quality checks with auto-retry, approvals, and skill-native execution. For most 3-6 step AI flows, describing what you want and generating YAML is faster than dragging nodes in n8n's editor.
Can I run flows without the desktop app?
Yes. c8c-workflow run flow.yaml runs the same engine headless. resume, rerun-from, and hil approve work from CLI too. Pipe it into CI, cron, or connect through OpenClaw for Telegram-triggered runs.
What happens when a step fails?
You can rerun from that specific step — the rest of the flow keeps its state. No need to restart from the beginning. If the evaluator triggers the failure, it auto-retries from the upstream step you configured.
Where are my flows stored?
Project flows live in {project}/.c8c/*.yaml. Global flows in ~/.c8c/chains/. Everything is local files, committable to git.
Is it really free?
Open source, MIT license. Runs locally. No account, no server, no fees. Works with your existing Claude Code, Codex, or OpenClaw subscription.
Development
npm run dev # Start Electron with hot reload
npm run build # Build for production
npm run canon:check # Check user-facing renderer copy against canon vocabulary
npm run test # Run all tests
npm run test:watch # Watch mode
npx tsc --noEmit # Type-check
Architecture
Electron app with three layers:
- Main (
src/main/) — Electron main process, IPC handlers, flow execution engine
- Preload (
src/preload/) — Context bridge exposing window.api
- Renderer (
src/renderer/) — React UI with list-based flow editor and runtime surfaces
Flows are directed graphs defined in YAML. The runtime expands the graph at execution time — splitter nodes create parallel branches, evaluators loop on failure. Each skill node spawns a fresh subprocess with clean context.
Stack: Electron 39, React 19, Tailwind CSS 3, Jotai, React Flow, Dagre, Vitest.
Contributing
c8c is early. The most valuable contributions right now are real flow YAML files, bug reports with reproduction steps, and documentation improvements. Code contributions are welcome too — check issues labeled good first issue. If unsure whether something is worth working on, open an issue first.
Community
License
MIT © 2026 c8c
