Clew describes what it's about to do, then does it — one visible step at a time. 16 providers, TUI + GUI, ReAct agents, MCP support, zero telemetry. Your code never leaves your machine.
No Python dependencies to install. No telemetry. Absolute privacy for local models, with the flexibility to plug in cloud APIs. Two frontends: Qt GUI and Textual TUI.
Local models run 100% on-device. No telemetry, no tracking, no cloud unless you want it.
Anthropic, OpenAI, Gemini, DeepSeek, Groq, xAI, z.ai, Nvidia NIM, or run Llama/Mistral/Qwen locally via Ollama or LM Studio. Plus 200+ via OpenRouter.
Qt-powered desktop app with real-time streaming, plus Textual TUI for terminal workflows. Switch seamlessly.
What matters when you're letting something else touch your code: what it plans, what it costs, and how it fits the way you work.
Before editing a file, Clew lays out its plan in plain language, then checks off each step. You always know what's about to change.
16 cloud providers or models on your own machine through Ollama, LM Studio, or Nvidia NIM. Switch the active model without leaving the project.
Token usage, cost per provider, and burn rate, tracked live. Set daily/monthly limits, enable predictable mode for deterministic token usage.
Git-aware filtering, relevance ranking, token budgeting, and live indexing. Clew reads your real project structure before proposing changes.
Built-in MCP client and manager. Connect external tools and knowledge bases — MCP tools appear alongside built-in agent tools.
Complexity-based provider and model selection with fallback chains. Automatically picks the best model for each task.
Drop .py plugins for custom providers and API routes. SKILL.md packages guide agent behavior with YAML frontmatter activation.
Built-in git tools: status, diff, stage, commit. Cross-chat memory with structured metadata so prior sessions inform new ones.
Python language server for autocomplete, hover, go-to-definition, and diagnostics — all running in a background thread.
LLM-based safety reviewer for risky tool calls. Three levels: off, dangerous-only, all. Shows 3-button modal (Approve / Reject / Use Fix) with proposed alternatives.
20+ pre-built templates across 9 categories (code, refactor, test, debug, document, review, deploy, office, learn). Run via /capabilities with placeholder substitution.
Second Opinion (Pro): ask a different model family to review risky actions. Verify last response: cross-family verifier checks correctness, safety, completeness after agent answers.
Add your own OpenAI-compatible endpoints via YAML config (~/.clew/providers.yaml) or drop Python plugins in ~/.clew/providers/. Hot-reload without restart.
A feature-by-feature look at how Clew stacks up against the alternatives.
| Feature | Clew | Cursor / Cline | Windsurf | Claude Desktop |
|---|---|---|---|---|
| Native desktop | Qt/PySide6 + Textual TUI, offline-capable | Web/Electron | Electron | Electron |
| Local-first privacy | Zero telemetry, 100% offline | Partial | Partial | Cloud-only |
| Autonomous agents | ReAct + JSON tool calling | Yes | Yes | Limited |
| Provider count | 16 native + 200+ via OpenRouter + custom | 5–10 | 3–5 | 1 (Anthropic) |
| MCP Protocol | Built-in client & manager | Plugin | No | Yes |
| Smart context | Git-aware, ranked, token-budgeted | Yes | Yes | Limited |
| Auto-Router | Complexity-based + fallbacks | No | No | No |
| Plugin system | Custom providers, routes, skills | No | No | No |
| Live token tracking | Real-time, per-provider quotas, budgets | Basic | No | No |
| Guardian safety | LLM reviewer, 3-level, fix suggestions | No | No | No |
| Cross-model verification | Second Opinion + Verify response | No | No | No |
| Capability catalog | 20+ templates, 9 categories | No | No | No |
| TUI frontend | Textual, full feature parity | No | No | No |
| Open source | MIT | Proprietary | Partial | Proprietary |
Add a key for the providers you use, or skip the cloud entirely and run everything on your own hardware. Custom providers supported via YAML or Python plugins.
...and 200+ more models via OpenRouter gateway
No account, no cloud dashboard to configure first.
Download the native app for macOS, Windows, or Linux from the latest GitHub release. Requires Python 3.11+. Or pip install -e . from source.
Paste an API key for any supported cloud provider, or point Clew at a local Ollama, LM Studio, or Nvidia NIM server. Auto-detection included. Add custom providers via YAML or Python plugins.
Open a project in the GUI (clew) or TUI (clew_tui) and tell Clew what you need in plain language. Review the plan, then let the agent read, write, and execute code.
No Docker, no Redis, no Node.js. Just Python + Qt. Native desktop with an embedded HTML5 frontend.
Clew is a solo project shaped with help from AI. It's rough in places and that's exactly why more hands would help — trying it and filing an issue is as valuable as a pull request.
Free and open source under the MIT License, for macOS, Windows, and Linux.