-
-
Notifications
You must be signed in to change notification settings - Fork 0
FAQ
An autonomous multi-agent framework that plans, codes, and validates software on your behalf, from a natural-language description.
Copilot and Cursor assist you while you code (autocomplete, chat). WorkPilot AI executes full tasks end to end: it plans, codes, tests, validates, fixes if needed, then proposes a merge. You stay in the loop via two checkpoints: spec approval and merge.
WorkPilot AI is open source under AGPL-3.0 — free to use. AI costs (Claude, OpenAI calls…) are on you and depend on your subscription / API key usage.
Yes, subject to AGPL-3.0:
- Internal use: no extra obligation
- Modified redistribution or exposed SaaS: your modified code must also be AGPL-3.0
A closed-source commercial license is available — contact the team.
It fully depends on the AI provider, chosen model, and task size. Rough guidance:
- SIMPLE task with Claude Haiku: a few cents
- STANDARD task with Claude Sonnet: €0.10–€1
- COMPLEX task with Claude Opus: €1–€10 or more
Use the Cost Estimator before launching, and watch the Analytics Dashboard afterwards.
Yes. WorkPilot AI supports OAuth for Claude and OpenAI — no paid API key needed if you already have an active subscription.
Recommended strategy:
| Task | Model |
|---|---|
| Spec creation | Sonnet |
| Complex planning | Opus |
| Coding | Sonnet |
| Simple QA Review | Haiku or Sonnet |
| QA Fix | Sonnet |
Use Arena Mode to compare and refine.
Yes, with Ollama (local models) + local Graphiti embeddings. Quality will be lower than frontier cloud models, but the option exists for max privacy.
Yes — files used in a task's context are sent to the AI provider. That's how LLMs work.
To limit:
- Filtering auto-redacts detected secrets
-
.env,.git,node_modules/are excluded by default - You can disable sending specific files via
.workpilotignore
Never in plaintext. Always in the native OS keychain:
- macOS: Keychain
- Windows: Credential Manager
- Linux: libsecret / kwallet
No. The dynamic allowlist systematically blocks destructive commands, and all filesystem operations are confined to the project directory.
- Changes live in an isolated worktree (main stays intact)
- You can abort anytime
- A Discard removes everything without touching main
- The 50 QA iterations cap prevents infinite loops
- Token budget is capped (
Settings → Security → Max budget)
Multiple safeguards:
- Spec Review: you approve before execution
- Human Review: you inspect the diff before merge
- Isolated worktree: main unchanged until merged
- Automatic QA: acceptance criteria validated by an independent agent
- AI Review: a second agent critiques the diff
- Tests required (via acceptance criteria)
Most common causes:
- Fuzzy acceptance criteria — rewrite them measurable
- Flaky existing tests that don't always pass
- Incomplete context — the agent didn't see a relevant file
- Conflicting criteria — two incompatible requirements
Check Agent Replay to see exactly where it's stuck.
In theory yes. That's why:
- Acceptance criteria are verified by an independent agent (QA Reviewer)
- A Workflow Logger traces everything
- You review the diff before merge
- AI Review offers a second AI pair of eyes
Tip: never blindly merge a task whose diff you don't understand.
The project must be an initialized Git repository. These frameworks are detected automatically (non-exhaustive): React, Vue, Angular, Svelte, Django, Flask, FastAPI, Express, NestJS, Next.js, Nuxt, Rust/Cargo, Go, Java/Maven, Java/Gradle.
For non-detected stacks, you can manually configure build and test commands in Settings → Project.
Yes. Each project has its own context, memory, and settings. Switch between projects via the list in the top-left.
Yes — up to 12 simultaneous agents (configurable). Each task = one isolated worktree, no interference.
WorkPilot AI is a standalone app. Work in parallel:
- VS Code to write yourself
- WorkPilot AI to delegate tasks to agents
An IDE integration is on the roadmap.
Yes, use the CLI: see CLI Usage. This is also the recommended approach for CI/CD.
- Read Contributing
- Browse issues labeled
good first issue - Join Discord
Yes — the code is modular. See apps/backend/integrations/ and Claude/OpenAI code as a reference. Open an issue to discuss before coding.
- Create a prompt at
apps/backend/prompts/<name>.md - Create an invocation via
create_client(agent_type="<name>") - Add the UI entry (button, palette) — frontend side
- Add i18n translations
Yes, on desktop — WorkPilot AI checks periodically and offers the update in a banner. You can disable in Settings → Advanced → Auto-update.
- Stable: recommended for everyone
- Beta: early features, potentially unstable
Change in Settings → Advanced → Channel.
Every release includes SHA256 and a VirusTotal scan published in the release comments. See Security.
No, a few MB per project typically. Periodic compaction (--compact) keeps size in check.
Yes: GRAPHITI_ENABLED=false or Settings → Memory → Disable. Each task then starts from zero.
No, each project has its own graph. This is deliberate: project A's conventions shouldn't pollute project B.
French and English officially. Other languages can be added by contribution — see Contributing.
Settings → Language — instant switch.
Yes, if the UI language is English AND the project allows it (code comments often stay in English per project convention). You can also specify in the spec: "error messages should be in English".
- See Troubleshooting
- Search GitHub Issues
- Ask in Discussions
- Join Discord
Getting started / Pour débuter
- 🏠 Home
- 📘 Introduction · fr
- ⚡ Installation · fr
- 🚀 First project · fr
- 🧠 Key concepts · fr
- ❓ FAQ · fr
Usage
- 🖥 User interface · fr
- 📝 Creating a spec · fr
- 🔁 Multi-agent pipeline · fr
- 🤖 Specialized agents · fr
- 🔌 Integrations · fr
- 💡 AI providers · fr
- 🧩 Memory system · fr
Advanced / Avancé
Community / Communauté