janus-agent 1.35.4

A self-improving AI agent for developers. Plain-text state, three gateways (CLI, Telegram, web), any OpenAI-compatible model.

A self-improving AI agent for developers. Plain-text state · three gateways · any OpenAI-compatible model.

---

What is Janus

Janus is a self-improving AI agent for developers. It learns from how you use it: noticing repeated patterns, proposing new skills you review and accept, evolving a memory of your preferences and projects over time.

State lives as plain-text markdown — auditable, hackable, syncable across devices. Three gateways (CLI, Telegram, Web) share one brain. Bring your own model from any OpenAI-compatible provider (Anthropic, OpenAI, OpenRouter, Ollama, llama.cpp, …).

Janus is not a SaaS product. It runs on your hardware. Your conversations, memory, and skills never leave your machine unless you explicitly send them somewhere.

---

60-second quickstart

# 1. Install (Linux / macOS — requires Python 3.10+)
pipx install janus-agent

# 2. Configure (interactive — picks up existing OPENAI_API_KEY etc.)
janus onboard

# 3. Chat
janus

That's it. You'll see a streaming chat with inline tool calls and a permission prompt for any side effect.

Want a deeper walkthrough? Four progressive tutorials in tutorials/ cover skills, memory, and MCP integration — each one screenful, ~10 minutes each.

Want to deploy on a VPS? scripts/install_services.sh installs systemd units for telegram + web + daemon with auto-restart on git pull.

---

How it works

flowchart LR
    User[You] --> CLI[CLI<br/>janus]
    User --> TG[Telegram bot<br/>janus telegram]
    User --> Web[Web UI<br/>janus web]

    CLI --> Loop
    TG --> Loop
    Web --> Loop

    Loop[executor.chat<br/>permission modes<br/>tool dispatch]

    Loop --> Tools[Tools<br/>fs · shell · web<br/>code_exec · MCP · …]
    Loop --> Skills[Skills<br/>~/.janus/skills/<br/>quarantined → trusted]
    Loop --> Memory[Memory<br/>~/.janus/memory/<br/>diff-and-approve]
    Loop --> Audit[Audit log<br/>~/.janus/log.jsonl<br/>every action recorded]

    Memory -.review diffs.-> User
    Skills -.review proposals.-> User

One agent loop. Three surfaces talk to it. State on disk you can read.

---

Why Janus

	Janus	Typical agent CLI
Model	Any OpenAI-compatible (Anthropic, OpenAI, OpenRouter, Ollama, …)	Provider-locked
State	Plain text under ~/.janus/ (cat, grep, git)	Opaque database
Skills	Markdown files; quarantined → user /promotes	None or auto-evolve
Memory	Diff-and-approve; user controls every save	Opaque embeddings
Audit	Every turn → ~/.janus/log.jsonl, FTS5-indexed	Best-effort
Gateways	CLI + Telegram + Web sharing one state	CLI only
Deploy	systemd / Docker / PyPI / git+URL	varies
Self-host	Always	Sometimes

---

Features

• Self-improving learning loop
  • Skills auto-proposed from repeated patterns; you review + promote.
  • Memory diffs proposed after each turn; you accept / reject / edit.
• Multi-surface
  • janus — terminal chat with streaming, tool inline, slash commands
  • janus telegram — bot gateway with inline plan-review keyboards
  • janus web — local FastAPI UI with cost charts, MCP browser, skills panel
  • janus daemon — runs scheduled agents (cron-style triggers)
• Plain-text everything
  • Memory: markdown cards under ~/.janus/memory/
  • Conversations: JSON files under ~/.janus/conversations/
  • Skills: markdown + YAML frontmatter under ~/.janus/skills/
  • Audit: JSONL log under ~/.janus/log.jsonl + audit.jsonl
• Permission modes — default / acceptEdits / plan / bypassPermissions. Set with /mode or JANUS_APPROVAL env.
• MCP integration — connect any stdio Model Context Protocol server (filesystem, git, sqlite, fetch, …). Browse the catalog with /mcp catalog.
• Production-ready
  • systemd integration: janus service install/enable/status
  • Reverse-proxy generators: janus web config caddy|nginx
  • Rate limiting on public endpoints (token-bucket per IP)
  • Backup / restore: janus backup / janus restore
  • Health endpoint: GET /api/health
  • Audit log: janus audit --action skill.promote
• Cross-device memory sync — janus sync push|pull (git-backed; bring your own remote)
• Webhooks — POST /api/webhook/<key> with HMAC-SHA256 fires the agent from any external service
• Image generation — built-in image_gen tool (DALL-E / Stable Diffusion via OpenAI-compatible endpoints)
• Cost tracking — /cost shows token usage + budget alerts at 50/80/100%

---

Install

Easiest — PyPI

pipx install janus-agent

For all gateways (Telegram + Web + headless browser tools):

pipx install 'janus-agent[all]'

One-line installer (auto-detects platform)

curl -sSL https://raw.githubusercontent.com/samalgotrader7-ops/janus/main/scripts/install.sh | sh

Docker

# Standalone web UI
docker run --rm -it -p 8765:8765 \
  -v janus-data:/root/.janus \
  --env-file .env \
  ghcr.io/samalgotrader7-ops/janus:latest web

# Three-service stack (web + telegram + daemon)
git clone https://github.com/samalgotrader7-ops/janus.git
cd janus
cp .env.example .env  # fill in your keys
docker compose up -d

From source

git clone https://github.com/samalgotrader7-ops/janus.git
cd janus
pipx install -e '.[all]'

---

Configure

Janus needs three env vars: an API key, an API base URL, and a model id.

The fastest path:

janus onboard

The interactive wizard walks you through provider selection (OpenRouter, OpenAI, Anthropic, Ollama, …) and picks up existing OPENAI_API_KEY / ANTHROPIC_API_KEY / OPENROUTER_API_KEY from your shell automatically.

Or set directly:

export JANUS_API_KEY="sk-..."
export JANUS_API_BASE="https://openrouter.ai/api/v1"
export JANUS_MODEL="anthropic/claude-haiku-4-5"

Persist these to ~/.janus/.env so they survive shell restarts.

---

Tutorials

Four progressive walkthroughs in tutorials/, each one screenful (~10 minutes each):

1. Hello, Janus — install, configure a model, first turn.
2. Your First Skill — write a skill, see it auto-load, /promote it.
3. Memory Loop — memory proposals, review, hygiene.
4. Connect MCP — configure a stdio MCP server, list its tools, call them through Janus.

---

Production deployment

For VPS / always-on setups, the systemd path is recommended:

git clone https://github.com/samalgotrader7-ops/janus.git /opt/janus
cd /opt/janus
# Set required env vars in your shell, then:
bash scripts/install_services.sh

What this gets you:

• Three systemd user-units: janus-telegram, janus-web, janus-daemon — each with auto-restart on failure.
• ~/.janus/.env written with chmod 600 from your shell env.
• loginctl enable-linger so units survive SSH logout.
• git config core.hooksPath scripts/git-hooks so git pull auto-restarts services when janus/*.py changes.
• Bypass with JANUS_NO_AUTO_RESTART=1 git pull.

For TLS, generate a reverse-proxy snippet:

janus web config caddy --domain janus.example.com >> /etc/caddy/Caddyfile
sudo systemctl reload caddy

(Both Caddy and nginx supported.)

---

Permission modes

Set with /mode <name> mid-session or JANUS_APPROVAL in .env:

mode	read	write	exec
default	allow	ask	ask
acceptEdits	allow	allow	ask
plan	allow	DENY	DENY
bypassPermissions	allow	allow	allow

plan mode is great for first sessions — Janus can read the codebase, propose a plan via the exit_plan_mode tool, and only start writing once you approve.

---

Where state lives

~/.janus/
├── memory/                  # markdown cards loaded into prompt
│   ├── MEMORY.md            # the index — always loaded
│   └── *.md                 # individual cards loaded on relevance
├── skills/<name>/SKILL.md   # markdown + YAML frontmatter
├── conversations/           # JSON per conversation; resumable
├── log.jsonl                # every turn + tool call + response
├── audit.jsonl              # mode changes / skill promotes / MCP connects
├── cost.jsonl               # token usage + USD spend
├── mcp/servers.json         # MCP server configs
├── triggers/*.json          # scheduled agents (run by `janus daemon`)
├── images/                  # generated images
├── backups/                 # `janus backup` output
└── .env                     # provider keys + config (chmod 600)

Everything is plain text. cat, grep, git diff, hand-edit at will.

---

Slash commands (cheatsheet)

/help                      list available commands
/mode <name>               change permission mode
/cost [--daily]            token usage + budget
/skills [show <name>]      list / inspect skills
/promote <name> <state>    promote skill (quarantined → trusted)
/memory review             review pending memory diffs
/memory consolidate        ask the model to dedupe memories
/mcp catalog               list configured MCP servers
/mcp connect <name>        connect to an MCP server
/resume [id]               resume a saved conversation
/clear                     start a fresh conversation
/why                       have model surface 2-3 alternative readings

---

Contributing

Janus is MIT-licensed and welcomes contributions. The codebase is plain Python with minimal dependencies — requests is the only required runtime dep; everything else is opt-in via extras.

git clone https://github.com/samalgotrader7-ops/janus.git
cd janus
pipx install -e '.[all,test]'
python -m pytest tests/ -q

Issues, ideas, and PRs welcome at github.com/samalgotrader7-ops/janus/issues.

---

License

MIT © Sam.