0agent

A persistent, learning AI agent that runs on your machine.

Runs a local daemon. Learns from every task. Remembers everything. Gets better over time.

npx 0agent@latest

What is this?

0agent is a CLI agent that runs as a background daemon on your machine. It executes real tasks — shell commands, file operations, web search, browser automation — using your API key, and learns from every outcome via a weighted knowledge graph.

Unlike chat-based AI tools, 0agent:

• Persists — runs in the background, remembers past sessions
• Learns — every task outcome updates edge weights in a graph; plan selection improves over time
• Executes — actually runs commands, writes files, searches the web, opens browsers
• Syncs — optionally backs up the knowledge graph to a private GitHub repo

Quick start

npx 0agent@latest

The wizard asks for:

1. LLM provider + API key (Anthropic, OpenAI, xAI, Gemini, or local Ollama)
2. GitHub repo for memory backup (optional, uses gh CLI if installed)
3. Workspace folder (where the agent creates files — default: ~/0agent-workspace)
4. Embedding provider (for semantic memory search)

After setup, the chat TUI opens automatically. No manual steps.

Usage

Interactive chat

0agent          # open chat (starts daemon if needed)
npx 0agent@latest

  0agent — anthropic/claude-sonnet-4-6
  Type a task, or /help for commands.

  › make a website for my coffee shop and deploy it locally
  › build a REST API in Go with auth
  › research my competitor's pricing and draft a response strategy

Type while the agent works — messages queue automatically and run one after another.

Slash skills

# Software engineering
0agent /review           # code review current branch
0agent /build            # run build, fix errors
0agent /qa               # generate and run tests
0agent /debug            # debug a failing test or error
0agent /refactor         # refactor a file or module
0agent /test-writer      # write unit tests
0agent /doc              # generate documentation

# Planning & strategy
0agent /office-hours "I want to build a payments feature"
0agent /plan-eng-review
0agent /plan-ceo-review
0agent /retro            # weekly retrospective
0agent /ship             # pre-release checklist

# Research
0agent /research "Acme Corp Series B"
0agent /security-audit

Scheduled tasks

  › /schedule add "run /retro" every Friday at 5pm
  › /schedule add "check the build" every day at 9am
  › /schedule list

Commands

| Command | Description | |---|---| | /model | Show or switch model | | /model add anthropic sk-ant-... | Add a provider API key | | /key anthropic sk-ant-... | Update a stored key | | /status | Daemon health + graph stats | | /skills | List available skills | | /schedule | Manage scheduled jobs | | /update | Update to latest version | | /graph | Open 3D knowledge graph | | /clear | Clear screen | | Ctrl+C | Cancel current task |

How it learns

Every task updates a weighted knowledge graph stored in ~/.0agent/graph.db.

Edge weights:  0.0 ──── 0.5 ──── 1.0
               bad    neutral   good

After each task:
  success → weight += 0.1 × learning_rate
  failure → weight -= 0.1 × learning_rate
  decay   → weight → 0.5 over time (forgetting)

After ~50 interactions, plan selection measurably improves. The graph also stores:

• Discovered facts: URLs, ports, file paths, API endpoints (via memory_write tool)
• Conversation history (last 8 exchanges injected as context)
• Identity + personality per entity

Memory sync

0agent can back up its knowledge graph to a private GitHub repo:

# Set up during init, or add manually to ~/.0agent/config.yaml:
github_memory:
  enabled: true
  token: ghp_...
  owner: your-username
  repo: 0agent-memory

• Pulls on daemon start
• Pushes every 30 minutes if there are changes
• Final push on daemon shutdown
• The same repo doubles as a GitHub Codespace template for browser sessions

What can the agent actually do?

| Capability | How | |---|---| | Run shell commands | shell_exec — bash, any CLI tool | | Read / write files | file_op — read, write, list, mkdir | | Search the web | web_search — DuckDuckGo, no API key needed | | Scrape pages | scrape_url — full page text, tables, links | | Open browser | browser_open — system Chrome or default OS browser | | Remember facts | memory_write — persists to knowledge graph | | Schedule tasks | Natural language cron via /schedule | | Self-heal | Detects runtime errors, proposes + applies patches |

Architecture

npx 0agent@latest
       │
       ▼
  ┌─────────────────────────────────────────────────────────┐
  │  CLI (bin/0agent.js + bin/chat.js)                      │
  │  • Init wizard  • Chat TUI  • Slash commands            │
  └───────────────────────┬─────────────────────────────────┘
                          │ HTTP + WebSocket
                          ▼
  ┌─────────────────────────────────────────────────────────┐
  │  Daemon (dist/daemon.mjs) — port 4200                   │
  │                                                         │
  │  SessionManager ── AgentExecutor ── LLMExecutor         │
  │       │                  │               │              │
  │       │            CapabilityRegistry    │              │
  │       │            • shell_exec          │              │
  │       │            • file_op             │              │
  │       │            • web_search          │              │
  │       │            • scrape_url          │              │
  │       │            • browser_open        │              │
  │       │            • memory_write        │              │
  │       │                                  │              │
  │  KnowledgeGraph ◄────── outcome feedback ┘              │
  │  (SQLite + HNSW)                                        │
  │       │                                                 │
  │  GitHubMemorySync ── SchedulerManager ── SelfHealLoop   │
  └─────────────────────────────────────────────────────────┘

Key packages:

| Package | Description | |---|---| | packages/core | Knowledge graph, inference engine, storage adapters | | packages/daemon | HTTP server, session manager, agent executor, capabilities | | bin/chat.js | Claude Code-style TUI with message queue, WS events, spinner | | bin/0agent.js | CLI entry point, init wizard, daemon lifecycle |

Configuration

~/.0agent/config.yaml — created by 0agent init, edit anytime:

llm_providers:
  - provider: anthropic
    model: claude-sonnet-4-6
    api_key: sk-ant-...        # never committed to git
    is_default: true

workspace:
  path: /Users/you/0agent-workspace   # agent creates files here

sandbox:
  backend: docker   # docker | podman | process | firecracker

github_memory:
  enabled: true
  token: ghp_...
  owner: your-username
  repo: 0agent-memory

embedding:
  provider: nomic-ollama   # nomic-ollama | openai | none
  model: nomic-embed-text
  dimensions: 768

Local development

git clone https://github.com/cadetmaze/0agentv1
cd 0agentv1
pnpm install
pnpm build

# Run init wizard
node bin/0agent.js init

# Or start daemon directly
node bin/0agent.js start
node bin/chat.js

# Bundle daemon into single file
node scripts/bundle.mjs

# Check status
node bin/0agent.js status
open http://localhost:4200    # 3D knowledge graph dashboard

Requirements:

• Node.js ≥ 20
• pnpm (npm install -g pnpm)
• API key for Anthropic, OpenAI, xAI, Gemini, or a local Ollama instance
• Docker (optional — enables sandboxed execution)

Roadmap

• [ ] Telegram bot interface
• [ ] MCP server support (connect to external tools)
• [ ] Team collaboration (shared graph, sync via GitHub)
• [ ] Mobile companion app
• [ ] Plugin SDK for custom capabilities

Contributing

Issues and PRs welcome. This is early-stage software — things break, APIs change.

1. Fork the repo
2. pnpm install && pnpm build
3. Make changes to packages/daemon/src/ or bin/
4. node scripts/bundle.mjs to rebuild the bundle
5. Test with node bin/0agent.js init
6. Submit a PR

License

Apache 2.0 — use it, fork it, build on it.