From one sentence
to your own AI-Native suite
WorkBench isn't a tool — it's a bench for building tools. You just speak your pain; it interrogates, researches, builds and delivers — a bespoke AI-native suite that fits your day, only yours. Here's Workbench-v0.1 and how its tools chain up.
Speak the pain
In plain words. No spec-writing needed.
Elicit the spec
Keeps asking + researches, distilling a build-ready spec.
Build it
Cheap models auto-code, gated + reviewed till it passes.
Plug abilities
Generate / publish etc., configured once, reused.
Your AI-Native suite
Deployed and running, only for you.
Turn "can't-articulate" into "build-ready"
Most people can't spell out their own pain. fde-copilot acts like a forward-deployed engineer: it keeps asking, researches what it can, asks only what only you know, and distills your words into six docs — Spec / Product / Features / Tech / Interactions / Gaps — so downstream can build with zero context.
- ① Create a client with a background (shared by all its projects)
- ② Create a project, name the deliverable + pick a type
- ③ Speak plainly; six docs + a readiness score grow live on the right
- ④ Once ready, commit and hand it to loop-engineer
Build the tool from the spec, autonomously
Takes the spec and dispatches cheap external models (GLM / Kimi / DeepSeek) to code autonomously in isolated git worktrees, running gate → cross-model review → rework → PR → merge until done. The orchestration is deterministic; the labor is cheap models; you only guard main at the end.
- ① plan breaks the spec into loop.json tasks with acceptance criteria
- ② Web console :4040: pick a job → ▶ Run
- ③ Watch live progress + per-task journal
- ④ Merges only after the gate + two-model review both pass
Atomic abilities: configure once, call forever
Building a tool calls base abilities behind the scenes: generate illustration / PPT / PDF, publish to WeChat / Xiaohongshu, fetch material for research… capability-packs wraps these into atomic abilities you configure once and reuse. Publishing needs your own account — fill it on the web page; credentials stay local, masked, never committed; missing ones are blocked, never fired with empty creds.
- ① Web :4141: enter platform accounts (local, masked, gitignored)
- ② Publish packs block until configured, with guidance
- ③ Agents invoke the abilities on demand while building
All open source. Run it fully yourself on a Mac Mini, or rent the public workbench for a few days, build your tool and take it with you — download, wipe or keep-encrypted anytime. You pay for convenience, reliability, expertise — not for access. The goal: ordinary people truly owning their own AI-native suite.
Architecture: the site lives on Cloudflare (public); the three tools are deployed independently and cross-link (a workbench switcher sits at the top of each). The "Open" buttons default to local ports 3939/4040/4141 — point them at your own domains after deploy.