crabagent 0.7.4

CrabAgent - AI Agent Platform with dual-mode (CLI + Serve)

🦀 CrabAgent

AI Team Command Center — Build a team of specialized AI agents that learn and improve over time. Delegate, parallelize, and watch them work in real-time from terminal or browser.

CrabAgent is a local-first AI agent platform. Run it from any project directory via CLI or browser. Your data stays local, your API keys are encrypted, and you pick any LLM provider.

---

Why CrabAgent

Unlike other agent platforms where agents are "temporary workers who forget everything," CrabAgent's agents learn and evolve:

Capability	What it means
🧠 Self-Evolving Agents	Agents auto-extract lessons from every task — rule engine catches patterns, LLM reflection analyzes strategies. The more you use them, the smarter they get.
🤖 AI Team	Custom agent profiles with per-agent tool whitelists and model overrides. Delegate, parallelize, or run multi-step pipelines.
📊 Agent Growth Tracking	View each agent's stats: task count, success rate, lessons learned, common task categories. ctrl+space agent_stats
⏱ Scheduled + Real-time	Agents run on cron schedules or react to @mentions. Real-time streaming of every agent's output.
🦀 Snapshots	Auto-snapshot files before changes. Roll back anytime without Git.
🔒 Local-first	All data stays on your machine. API keys encrypted at rest. No telemetry.

---

Quick Start

pip install 'crabagent[serve]'

crabagent init

# TUI — interactive REPL with slash commands
crabagent

# TUI (legacy single-panel)
crabagent --old

# Web UI
crabagent --serve          # → http://localhost:5210
                           # Default login: admin / xcl1989

# Single-query CLI
crabagent "organize this directory"
crabagent -p deepseek -m deepseek-chat "write a Python script"

---

Self-Evolving Agents

This is CrabAgent's core differentiator. Agents don't just execute tasks — they learn from every execution.

How it works

Sub-agent completes task
    │
    ├─ Rule Engine (instant)
    │   └─ High iterations (>80% max) → "Decompose complex tasks, use fewer tools per step"
    │
    └─ LLM Reflection (~1-3s)
        ├─ Extracts concrete, actionable insights:
        │   "When searching Chinese news, use English keywords on DuckDuckGo for better results"
        │   "For error-prone sites, prefer web_scrape with direct URLs over web_search"
        ├─ Auto-filters generic/noise responses ("completed in X steps")
        ├─ Also learns from failures: captures what went wrong and how to avoid it
        └─ Source: llm

Knowledge persistence

• Team Knowledge: Tech stack, architecture decisions, user preferences — auto-injected into every session
• Agent Lessons: Per-agent concrete insights grouped by category (Pitfalls / What Worked) — loaded before similar tasks
• Task Records: Every execution logged (success, elapsed time, tokens, iterations)

Tracking growth

# TUI
/agent_stats coder
# → 总任务: 23  成功率: 91%  平均耗时: 14s
# → lessons: 18 (规则: 3, LLM: 15)

# Web UI
# → Agent Team → Learning Stats: click agent name to see stats + all lessons

---

## AI Team

### Built-in agents

| Agent | Role | Best For |
|-------|------|----------|
| 🔍 Researcher | Web research | Search, browse, data collection |
| 📊 Analyst | Data analysis | Comparison, pattern detection, reports |
| 💻 Coder | Code expert | Write, review, debug, refactor |
| 📝 Writer | Content writer | Write, edit, translate, format |

### Delegation

- `@researcher find competitor pricing` — @mention auto-delegates
- Click an agent from the toolbar to insert a mention
- `/delegate` command for interactive agent selection
- `delegate_parallel` runs multiple agents simultaneously
- `run_pipeline` chains agents with dependencies

### Session Agent Switching

Switch your current agent identity mid-session without losing conversation history:

```bash
# TUI
/agent                  # Popup menu: select from 5 agents
/agent researcher       # Direct switch
/agent default          # Back to all tools

# Web API
POST /api/sessions/{id}/agent  {"agent": "researcher"}

• Each agent has different tool sets (researcher gets web tools, coder gets bash+edit, etc.)
• System prompt stays unchanged — LLM KV cache preserved across switches
• All messages are tagged with agent info for history tracking
• Model auto-switches if the agent profile specifies one
• Status bar shows current agent: [deepseek/chat → researcher] Msgs:5 Tok:1234

Real-time monitoring

• 🟣 Running — live step count and timer
• 🟢 Done — elapsed time, tokens, iterations
• 🔴 Error — error summary
• Web: right-side Task Board with split-pane result comparison

---

More Features

🖼️ Multimodal

Paste, upload, or drag images. Auto-detects vision models.

🌐 Browser Automation

pip install 'crabagent[browser]' + playwright install chromium

> Open https://news.ycombinator.com and show top 5 stories
> Search "Python async" on Google

🔌 MCP Client

Connect external MCP servers (stdio + HTTP). Tools auto-discover.

📋 Scheduled Tasks

> Remind me every day at 11:00 to drink water
> Check product page every 30 minutes, notify if price drops

🦀 Snapshots

Auto-snapshot before file changes. Roll back with /molt rollback <id>.

🔧 Custom Plugins

Drop a .py file in .crabagent/tools/:

name = "hello"
description = "Say hello"
parameters = {"type": "object", "properties": {"name": {"type": "string"}}, "required": ["name"]}
requires_permission = False

def run(name: str) -> str:
    return f"Hello, {name}!"

Or let agents create tools themselves — Your agent can write and register custom tools during a session. Tell it what you need and it will generate, validate, and save the tool:

> Create a tool that parses CSV files and extracts a column
> Create a tool to fetch weather for a city

Tools are saved to .crabagent/tools/, auto-registered, and persist across sessions. The agent remembers its created tools via team memory.

---

CLI / TUI Commands

Command	Description
/exit, /quit	Exit
/help	Show help
/clear	Clear context
/model [name]	Switch model
/models	List models
/provider [cmd]	Manage providers
/sessions / /session [id]	List / load sessions
/new	New conversation
/agents [cmd]	Agent team management
/agent [name]	Switch current agent
/agent_stats <name>	Agent growth stats
/delegate [@agent] [task]	Delegate task
/memory [list|search|clear]	Team memory
/skills / /skill <name>	List / show skills
/molt [cmd]	Snapshots
/todo [cmd]	Task list
/export	Export to Markdown
/image <path> [msg]	Send image
/runs [agent]	View agent run history
/abort	Abort current agent (Ctrl+C)

---

Configuration

Variable	Default	Description
CRAB_DB_URL	sqlite+aiosqlite:///./crabagent.db	Database URL
CRAB_JWT_SECRET	Auto-generated	JWT signing key
CRAB_SERVE_HOST	0.0.0.0	Serve host
CRAB_SERVE_PORT	5210	Serve port
CRAB_MAX_ITERATIONS	50	Max agent iterations
CRAB_MAX_TOKENS	4096	Max response tokens
CRAB_BROWSER_HEADLESS	true	Browser headless mode
CRAB_WEB_PROXY	(empty)	HTTP proxy for web_search & web_scrape

v0.7.4 Highlights

• 🔄 Session Agent Switching — Switch agent identity mid-session with /agent (TUI) or POST /api/sessions/{id}/agent (API). Each agent has different tool sets, and messages are tagged with agent info for history tracking.
• 🛠️ Agent-Created Custom Tools — Agents can now write and register their own reusable tools via create_tool/update_tool/delete_tool. Code is validated, saved to .crabagent/tools/, registered immediately, and auto-loads across sessions.
• 🐛 TUI Queue & History Fixes — Fixed race condition where queued inputs were sent before previous rendering completed. Fixed message ordering in DB when loading sessions with queued messages via persistence flush improvements.
• 🔤 TUI CJK & Thinking Fixes — Fixed CJK character rendering freeze in dual-panel TUI. Fixed thinking text display bugs (off-by-one, cache miss on content update, prefix loss on flush).

v0.7.2 Highlights

• 🖥️ Dual-Panel TUI — New prompt_toolkit-based full-screen TUI: scrollable output panel (mouse wheel + PageUp/Down/Home/End), persistent input area that auto-grows with content, and real-time status bar. Default mode (crabagent), use --old for legacy TUI.
• 🖱️ Mouse Text Selection — Hold Shift + mouse drag to select text in the output area. Ctrl+C copies to clipboard (macOS pbcopy / Linux xclip).
• 💬 Interactive Popup Menus — /model, /sessions, /provider now show scrolling selection popups with arrow key navigation, instead of printing long lists.
• 🧠 Streaming Thinking — Agent reasoning (THINKING_DELTA) streams in real-time to the output panel, dim italic style.
• 💡 Completions Menu — Slash command autocomplete appears as a floating completions menu above the input area.

v0.7.1 Highlights

• 📊 Pipeline Dashboard — Real-time pipeline visualization: see active pipelines with step progress, agent cards with running counts, and growth charts. History pipelines auto-collapsed.
• 🔄 AgentRun Persistence — New agent_runs table tracks every agent/pipeline execution with full metadata (tool calls, elapsed time, tokens, iterations). API endpoints for run history and per-agent growth stats.
• 🐛 Streaming Fix — TEXT_DELTA and THINKING_DELTA events are no longer throttled/dropped by SSE forwarder. TEXT_DONE handler now uses full text from backend to ensure complete message display.
• 🛠 Tool Display Fix — delegate_parallel arguments with nested objects no longer show [object Object] in the UI.
• 📡 RunRecorder — EventBus subscriber that creates agent_runs records for pipeline, main, and sub-agent executions in real-time.

v0.7.0 Highlights

• 🧠 Learning quality overhaul — LLM reflection now extracts actionable insights (tool tricks, pitfalls, domain tips), no more "completed in X steps" noise. Failure learning added — agents learn from mistakes too.
• 🌐 Web proxy support — CRAB_WEB_PROXY=http://127.0.0.1:7890 for web_search/web_scrape (critical for users behind firewalls).
• 📊 Learning Dashboard — View each agent's task stats and past lessons directly in the Web UI Agent Team panel.
• 📡 Sub-agent persistence — Completed sub-agents stay visible in the Dashboard for 30 minutes after finishing.

---

Installation

pip install 'crabagent[serve]'          # Web UI + API
pip install 'crabagent[browser]'        # Browser automation
pip install 'crabagent[dev]'            # Testing + linting

# Development
make install            # Build frontend + install (editable)
ruff check src/ tests/  # Lint
ruff format src/ tests/ # Format
pytest                   # Run tests

---

Project Structure

CrabAgent/
├── src/crabagent/
│   ├── cli/           # CLI entrypoint + TUI
│   ├── core/agent/    # Agent loop, tools, compression, agents
│   ├── core/mcp/      # MCP client manager
│   └── serve/         # FastAPI + API + scheduler
├── frontend/          # React SPA
└── crabagent.db       # SQLite database

---

License

GNU Affero General Public License v3 (AGPLv3) for non-commercial use. Commercial use requires a separate license. Contact the author.

See LICENSE.