An agent for coding and everyday work that runs on the models you choose: your ChatGPT plan, any OpenAI-compatible API, or a model on your own GPU.
Terminal, desktop and web on one local engine. It matches stock Hermes on pass rate while using under a third of the prompt.
# macOS · Linux
curl -fsSL https://raw.githubusercontent.com/Picom-code/pico/main/installer/install.sh | sh
# Windows (PowerShell)
irm https://raw.githubusercontent.com/Picom-code/pico/main/installer/install.ps1 | iex
Then run pico. The first launch walks you through choosing a model. Updates are offered, never applied automatically. Installer details are in installer/README.md.
Why pico
- Lean by default. It passes the same 60/60 eval as stock Hermes with 72% fewer prompt tokens per task and 34% less wall time. The first turn sends about 2.0K tokens instead of about 9.1K.
- A real agent, not a chat box. Minimal harnesses like Pi drop features that pico keeps from Hermes: memory, skills, crons, subagents and background replies.
- Your models, your machine. Sign in with ChatGPT, paste any OpenAI-compatible endpoint, or let pico scan your hardware and pick a local model that fits. Your data stays in
~/.pico.
- Pleasant to use. You can steer mid-reply, roll back with
/undo (file checkpoints included), and accept tab completions from ghost text. The terminal UI was tuned frame by frame. You get the cat, not emoji fireworks.
Benchmarks
The suite has 30 tasks (coding, ops, everyday tasks, memory, skills, long context), each run twice, on deepseek-v4.1-flash. Full method: docs/EVAL.md · docs/LEAN.md.
| metric |
stock Hermes |
pico lean |
Δ |
| pass rate |
60 / 60 |
60 / 60 |
±0 |
| prompt tokens / task (median) |
57.9K |
16.4K |
−72% |
| wall time / task (median) |
43.5s |
28.6s |
−34% |
| first-turn prompt |
~9.1K |
~2.0K |
−78% |
Lean also holds its pass rate as filler context grows from +0 to +32K tokens, with no cliff (context curve).
Features
|
In the chat
- Steer: press
enter mid-reply and your message lands while the agent keeps working
/plan drafts a plan, then waits for your go
/goal pins a goal banner and checklist
/undo rolls back the last turn, files included
/yolo runs commands without asking (toggle)
@file attachments, tab completion with ghost text
- Live tool tree, real diffs,
ctrl+o to expand
|
Around the chat
/loop 10m "check CI" repeats on a timer or until the task is done
/cron takes plain English and delivers results to /inbox
/usage shows tokens, time and tok/s per model over 30 days
/model picker with default effort, shift+tab per chat
/lean slims your profile and shows the measured speedup
- Auto-compaction (summary or trim),
/compact [focus]
- Subagents panel: parallel helpers, each with its own model and budget
|
|
Make it yours
/settings tabs in the style of Claude Code: Status · Config · Usage · Look & feel
- 6 accents, 8 spinners, 4 reply styles, ASCII fallback
- Verbosity from quiet to full, set separately for reasoning and tools
|
Setup and other surfaces
- ChatGPT login, or any OpenAI-compatible API (presets plus a live test)
- A local hardware scan picks a model that fits your GPU or Mac
- Desktop app (Tauri, 4.3 MB) and a web app on the same engine
- Phone pairing over Tailscale (iOS remote planned)
|

/settings |

/model |
Keys and commands cheat sheet
| key |
does |
enter / shift+enter |
send / new line (send while busy = steer) |
esc |
stop the current reply |
shift+tab |
cycle effort for this chat (low · med · high) |
ctrl+o |
expand or collapse thinking and tool output |
tab |
accept the ghost-text completion |
@ |
attach a file (fuzzy search) |
ctrl+l |
redraw |
ctrl+u / ctrl+k / ctrl+w |
kill line / to end / word |
| command |
does |
/model [id] |
switch model, set default effort |
/setup · /login |
ChatGPT, any API, or local |
/new · /resume · /sessions · /search |
chats |
/plan <task> · /goal [text] |
plan first / pin a goal |
/loop <5m|until-done> … · /cron · /inbox |
recurring and scheduled work |
/undo · /yolo [on|off] |
checkpoints / approvals |
/compact [focus] · /lean · /usage |
context and cost |
/agents · /memory · /skills · /connect |
helpers, memory, skills, apps |
/settings · /theme · /verbose |
look and feel |
/copy · /clear · /help · /quit |
the rest |
Scripting: pico -p "prompt" prints one answer and exits (--json emits events). Examples: git diff | pico -p "review this", pico doctor.
Desktop
 new chat |
 model picker |
 settings |
How it fits together
flowchart LR
T["pico terminal<br/>TypeScript · Bun"] -- "HTTP + SSE" --> E
D["Pico.app<br/>Tauri"] --> W["web app<br/>Preact"]
W -- "HTTP + SSE" --> E
P["phone (planned)"] -. Tailscale .-> E
E["pico engine<br/>Python · 127.0.0.1:7878"] --> H["Hermes Agent<br/>library, pinned"]
H --> M["models<br/>ChatGPT · OpenAI-compatible · local llama.cpp"]
E --- S[("~/.pico<br/>chats · memory · skills · crons")]
Loading
| dir |
what |
engine/ |
Python engine: sessions, lean harness, compaction, crons, hardware scan, setup. API reference: docs/API.md |
cli/ |
Terminal app (TypeScript, Bun single binary) |
web/ · desktop/ |
Preact web app and the Tauri shell around it |
installer/ |
install.sh / install.ps1: uv-managed Python, pinned Hermes, launcher |
eval/ |
Harness eval suite and context-curve runner |
recipes/ |
Local model catalog used by the hardware scan |
Roadmap
- Engine on Hermes Agent, lean harness, eval suite
- Terminal app: steer, undo, plan, goal, loop, cron, usage, settings
- Web app and desktop app (Tauri)
- One-line installer (macOS · Linux · Windows)
- Signed desktop builds and prebuilt binaries for every platform
- iOS remote over Tailscale
- Local voice, image and video
- Catalog agent that offers faster local models as they ship
Contributing
PRs are welcome. CONTRIBUTING.md covers dev setup, tests and the eval. Report security issues by following SECURITY.md.
Credits
pico is built on Hermes Agent by Nous Research (MIT). The agent loop, tools, memory and skills come from Hermes. The lean harness, engine, apps and installer were written for pico.
License
MIT © 2026 Patrick Moorhead
( ^.^ ) tests first, then snacks.