Your AI agent, your stack
Paw is a personal AI agent framework: multi-provider, MCP-native, with a live canvas, on-demand skills, vector memory, and a web UI. Bring Claude, GPT, Gemini, or Ollama. Swap them anytime.
Bun runtime. MCP-native. Skills on demand. Open source.
curl -fsSL https://paw.sh/install | bashBuilding AI agents is hard composable
Every framework picks its own model, its own tools, its own UI. Paw gives you the pieces without the lock-in.
What Exists Today
- Vendor lock-in (one provider, one model)
- Cloud-only agent platforms (no local-first)
- Glue code in every project to wire tools together
- Monolithic frameworks that do too much
- Hard-coded skills you can't add or remove
What Paw Does
- Multi-provider routing: Claude, GPT, Gemini, Ollama
- Local-first kernel with a live web UI
- MCP-native — add any tool with one line
- Skills activate on demand, no restart required
- Cron, webhooks, and a canvas for visual workflows
Built for shipping
Everything you need to run an AI agent in production, nothing you don't.
Multi-Provider Routing
Claude, OpenAI, Gemini, and Ollama in one config. Switch models per request, fall back automatically on errors.
MCP-Native
First-class Model Context Protocol support. Add any MCP server with one line — tools stream in instantly.
On-Demand Skills
Skills are plugins that activate when needed. Write TypeScript, drop it in, the agent picks it up.
Vector Memory
Long-term memory with semantic recall. The agent remembers your preferences, your codebase, your past work.
Live Canvas
A web UI for visual workflows, canvas pages, and human-in-the-loop approvals. Edit and ship in one place.
Cron + Webhooks
Schedule agent runs with cron. Trigger them from external services with webhooks. Proactive triggers built in.
Tool SDK
Define a tool once, expose it via MCP, REST, and the canvas. Type-safe, with permission gates and audit logs.
Bun Runtime
Built on Bun for fast startup, native TypeScript, and a tiny memory footprint. One binary, zero ceremony.
Web UI + REST API
Ship a UI and an API from the same config. No separate frontend/backend build pipelines to maintain.
Get started in one command
Three install paths. Pick the one that fits your workflow.
One command. Works on macOS, Linux, and WSL.
curl -fsSL https://paw.sh/install | bashpaw startopen http://localhost:3030Plugs into everything
Swap models. Wire tools. Drop in any MCP server. Paw meets your stack where it lives.
Models
Anthropic Claude
ModelSonnet, Opus, Haiku
OpenAI
ModelGPT-4o, o1, o3-mini
Google Gemini
ModelPro, Flash, Nano
Ollama
ModelLocal models, no API key
Tools & Integrations
GitHub
Repos, issues, PRs
Slack
Channels, DMs, reactions
Gmail
Read, draft, send
Notion
Pages, databases
Postgres
Query, schema, migrations
Web Search
SerpAPI, Brave, Tavily
Discord
Servers, channels
AWS S3
Buckets, signed URLs
Any MCP
Drop-in via stdio
Missing a tool? Write a 30-line MCP server and Paw picks it up automatically. No framework changes required.
Four-layer architecture
Small surface area. Clean boundaries. Every layer swappable.
Provider Matrix
All providers implement the same interface. Switch by config, not refactor.