Product Docs Blog GitHub ↗
v3 — Now with Swarm Intelligence
Bamboo — PandaClaw's space panda

PandaClaw

Slow is smooth. Smooth is perfect.

A deliberate, reasoning-first personal AI that plans before it acts.
Swarm workers, git-backed rollbacks, episodic memory — all from your terminal or Telegram.

Get PandaClaw View on GitHub
Bun ≥1.3.3 TypeScript v5 Groq Llama 3 DeepSeek R1 MIT License
Quick Start
macOS · Linux · Windows 
# Installs Bun, clones, registers binary. You're welcome. 🐼
$ curl -fsSL https://pandaclaw.vercel.app/install.sh | bash
# Installs Bun, clones, registers binary (PowerShell)
PS> irm https://pandaclaw.vercel.app/install.ps1 | iex
# Global install using Bun (native runtime)
$ bun install -g pandaclaw
# or using npm
$ npm install -g pandaclaw
# Clone and link globally for development
$ git clone https://github.com/senapati484/pandaclaw.git
$ cd pandaclaw && bun install && npm install -g .

What People Say

View all

"I've been running PandaClaw on my laptop for a week now. It feels like the shift from Windows to Linux 20 years ago. You're fully in control..."

openpke
openpke @openpke

"Today was one of those days I ran to my computer after daycare drop-off. PandaClaw had already handled three tasks I left mid-thought the night before."

kyle zantos
kyle zantos @kylezantos

"PandaClaw is absolute magic. A reasoning-first Swarm agent running locally on Bun with episodic memory and full sandboxed rollbacks!"

Zilla
Zilla @hey_zilla
Core Architecture

Built different.
Thinks before it acts.

Six interconnected systems that make PandaClaw a truly deliberate AI agent.

🔒

Git-Backed Rollbacks

Every file mutation runs on a temp branch. Decline and the workspace snaps back instantly.

🐝

Coordinator-Worker Swarm

Tasks decompose into a dependency tree, then run concurrently across four specialised workers.

🧠

Episodic Memory Graph

Facts and patterns from raw logs stored as triplets in a persistent knowledge graph — recalled in under 1ms.

🎨

Local Visual Canvas

Glassmorphic dashboard at localhost:18789 with real-time chat, vision overlays, and interactive diff approval.

👁️

4-Stage Vision Pipeline

Perceive → Locate → Reason → Act. Sees your screen, identifies elements, and takes targeted actions.

⚡

Unified Event Gateway

One adapter layer routes text and screen images through the same reasoning brain.

✈️
Telegram Bot Mobile companion
⌨️
CLI Terminal Shell execution
🌐
WebChat Dashboard localhost:18789
Reactor Loop

Six steps from prompt
to perfect output.

PandaClaw never rushes. Every task moves through a strict reasoning pipeline before a single file is touched.

01

Observe

Index the codebase, load episodic memory, and hydrate the Knowledge Graph with relevant facts.

02

Reason

Evaluate goal status and allocate R1 thought budget using DeepSeek's chain-of-thought traces.

03

Plan

Generate a topological task graph, establish critical path, and assign workers by specialisation.

04

Execute

Low-risk tasks auto-run. High-risk mutations (deletion, shell) enter a Git sandbox requiring your approval.

05

Validate

The verifier worker checks syntax, runs tests, and applies critique-correction loops before committing.

06

Reflect

Record success or failure, update the episodic graph, and make every future task smarter.

Risk Safety Model

Safe defaults. You're always in control.

Level Mutation Type Behaviour
Low Directory creation, file edits < 50 KB Auto-execute & validate
Medium Large file edits, multiple mutations Auto-execute up to threshold
High Deletion, shell execution, config changes Git sandbox + interactive approval
Real-World Use

What people build
with PandaClaw.

From automation to coding workflows, PandaClaw adapts to how you work.

🤖

Developer Automation

Automate code reviews, run tests, fix issues, and create PRs directly from chat.

📊

Personal Dashboard

Daily briefings, calendar summaries, and reminders through one assistant.

🎯

Business Process

Handle expense reports, document drafting, and multi-step workflows from Telegram.

Quick Start

Up and running
in one command.

Three ways to install PandaClaw — pick the one that fits your workflow.

Recommended
01

One-liner Script

Automated installer — checks/installs Bun, clones the repo, and registers the global binary in one shot.

BASH 
# macOS & Linux
$ curl -fsSL https://pandaclaw.vercel.app/install.sh | bash

# Windows (PowerShell)
PS> irm https://pandaclaw.vercel.app/install.ps1 | iex
02

npm / Bun

Install from the public npm registry using your preferred package manager.

BASH 
# Using Bun (native runtime)
$ bun install -g pandaclaw

# Using npm
$ npm install -g pandaclaw
03

Manual Clone

For local development or contributing to PandaClaw.

BASH 
# Clone repository
$ git clone https://github.com/senapati484/pandaclaw.git
$ cd pandaclaw && bun install

After installing, run the setup wizard

1
Configure pandaclaw setup
→
2
Wake up pandaclaw wakeup
→
3
Open canvas localhost:18789
Configuration

Your keys.
Your machine.
Your agent.

PandaClaw connects to Groq, OpenRouter, and NVIDIA NIM. Run the wizard or drop a config.json. Environment variables always win.

  • 🦙
    Groq Llama 3 Fast chat & memory consolidation
  • 🤔
    DeepSeek R1 via OpenRouter Deep reasoning & R1 compiler
  • 👁️
    NVIDIA NIM 4-stage vision pipeline
  • 📱
    Telegram Shared bot token — no BotFather needed
Full config reference
config.json 
{ 
  "providers": {
    "groq": {
      "api_key": "YOUR_GROQ_KEY"
    },
    "openrouter": {
      "api_key": "YOUR_OR_KEY"
    },
    "nvidia_nim": {
      "api_key": "YOUR_NIM_KEY"
    },
    "ollama": {
      "api_base": "http://127.0.0.1:11434/v1"
    }
  },
  "telegram": {
    "allowed_users": [12345678]
  }
}