poseidon-ai 0.14.1

Open-source agent harness that opens as a chat in your browser. Watch it work, approve what matters, see what it costs.

🔱 Poseidon

An open-source agent harness that opens as a chat in your browser. Watch it work. Approve what matters. See what it costs.

Most agent harnesses live in a terminal. Poseidon doesn't. One command installs it, one command opens a chat on localhost — with a live workspace pane showing every file it touches, every command it runs, and a running meter of what it's spending.

pipx install poseidon-ai
poseidon ~/my-project

Your browser opens. You talk. It works — visibly.

Why another harness?

Three things the others don't do:

1. Watch it work. Split-pane UI: chat on the left, live workspace on the right — activity stream, file browser, every tool call visible as it happens. No claims in prose you can't verify; you see the hands moving.
2. The trust dial. File writes and shell commands pause and ask, inline in chat, with a preview. Hit Always allow and Poseidon earns that permission permanently. Autonomy is granted, not assumed — and it accumulates.
3. It knows what it costs. A live cost meter — per session, per model, tokens in/out — built into the header. No more waking up to a burned API budget.

And it's properly agentic, not just a chat with tools:

• Plans first — multi-step work shows up as a live checklist in the workspace
• Subagents — big chunks get delegated to parallel subagents with their own context
• Schedules — "check this every morning" becomes a real scheduled run (approval-gated; unattended runs can only do what an Always allow rule already covers)
• Memory — durable facts persist across sessions as plain markdown in ~/.poseidon/memory/. Your agent's memory is your files: open it, edit it, delete it.

And because you stare at it all day: skins. Daylight waves (default), Trek Wars (starfield + hyperjumps), Ukiyo-e (living woodblock waves), and Wasteland (dunes, dust, the occasional tumbleweed). All hand-drawn vector — sharp at any size.

Works with any model

Any OpenAI-compatible endpoint:

• Ollama — fully local, zero API key, free. Nous Hermes runs great.
• OpenAI, Groq, or any custom endpoint
• Bonito — one key for every provider, with failover and team-level cost tracking
• Your ChatGPT subscription (experimental) — "Sign in with ChatGPT" in Settings opens the normal ChatGPT sign-in in your browser (the same OAuth flow codex login uses), no API key and no account settings needed. On a remote/headless box use the device-code fallback instead — note that's an OpenAI beta that must be enabled first (ChatGPT → Settings → Security → "Device code login"; workspace accounts: your admin turns it on). New and lightly tested; if it misbehaves, any of the endpoints above work.

Each preset carries the model's real context window, so long sessions auto-summarize before the model's limit instead of erroring — the meter in the toolbar shows how full the session is.

Keys are stored in ~/.poseidon/config.json on your machine and sent nowhere else.

Project instructions

Drop an AGENTS.md in your working directory and Poseidon reads it — same convention as every other harness. No new file formats to learn.

Status

v0.5: the core loop (chat, file tools, shell with approvals, web fetch, cost meter) plus task planning, parallel subagents, background tasks, scheduled runs, team projects (profiles, shared sessions, handoff notes, status), automatic context compaction, checkpoints (auto + rewind), persistent project memory, and a live Pipeline diagram of everything running with drill-down timelines. Architecture in ARCHITECTURE.md; roadmap in PLAN.md.

Development

git clone https://github.com/ShabariRepo/poseidon
cd poseidon
python3 -m venv .venv && .venv/bin/pip install -e .
.venv/bin/poseidon --no-browser

License

MIT — made by Bonito AI Labs.