Do I need an API key?

No. The model is served by the Animica miner network through a public keyless endpoint, so pip install animica and animica chat is the whole setup. A key is accepted if you have one, but nothing requires it.

A turn typically takes 30 to 150 seconds, because inference comes from volunteer miners rather than a datacentre. It streams as it writes and re-submits to another miner when one does not answer. If you need sub-second replies, this is not the right tool.

Can it change my files without asking?

Not in the default mode. manual mode asks before writing a file, deleting, running a shell command or fetching a URL. plan mode refuses those outright and will not even offer a prompt. auto-edit applies edits but still asks before shell and delete. auto approves everything and requires typed confirmation to switch into mid-session.

What does a paid plan add?

Unlimited agentic tasks instead of ten a day, fifty tool-call rounds per task instead of ten, and eight parallel agents in /swarm instead of two. Chat itself is unlimited on the free tier.

Terminal · animica chat

An agentic coding assistant with no API key.

animica chat reads and edits the files in your working directory, runs commands, and asks before anything with a consequence — the way Claude Code and Codex do. There is no account to make, no key to paste, and no GPU to own.

# a venv keeps it out of your system Python
python3 -m venv .venv && source .venv/bin/activate
pip install animica
animica chat

On Windows the activate line is .venv\Scripts\activate. Prefer one command? curl -fsSL https://animica.dev/install.sh | sh — and you can read it first; it needs no sudo.

The honest part, before you install anything. Inference comes from volunteer miners, not a datacentre, so a turn usually takes 30–150 seconds — and right now very few machines serve chat, so the first request after an idle period is the slowest. The CLI streams as it writes and re-submits to another miner when one does not answer, but it will never feel like a hosted API. If you need sub-second replies, use something else and come back when the network is bigger.

What it actually does

Tools are live on every message — there is no separate agent mode. Ask what a function does and ask for it to be fixed in the same session.

Reads, greps and globs your files; writes, edits and moves them; runs shell commands and Python.
Streams its answer as it arrives, so a long wait shows work rather than a spinner.
Reads an AGENTS.md (or CLAUDE.md, .animica/instructions.md) from the directory or any parent, and says in the banner which file it is following.
Takes a pipe: pytest 2>&1 | animica chat "why is this failing?"
Sessions are plain JSON on your machine. animica chat -c continues the last one.
Arrow-key history, and /swarm for parallel agents.

# one shot, then exit
animica chat "find the N+1 query in this repo and fix it"

# start with edits pre-approved, shell still asks
animica chat -p auto-edit

You decide what it may touch

Four approval modes, switchable mid-session with /mode or set at start with --permission-mode. The mode is always on screen, because an agent that auto-approves shell commands without saying so is indistinguishable from a careful one until something is already gone.

plan

Reads only. Writes, shell and network are refused — it will not even ask.

manual

The default. Asks before writing a file, deleting, running a shell command or fetching a URL.

auto-edit

Edits apply as it goes; shell and delete still ask. What most coding sessions want.

auto

Everything, unattended. Switching into it mid-session makes you type auto to confirm.

Modes are built on a tool category map rather than one safe/unsafe bit, which is what lets “edit freely, ask before shell” exist as a real setting. A tool nobody has categorised is treated as executable, so it stays gated.

`/swarm` — and a reviewer that tries to prove it wrong

Type /swarm audit this codebase for auth bugs inside a session. It splits the work into parallel subtasks, runs them together, then has a separate agent try to refute each result and merges only what survives. A review that comes back inconclusive counts as refuted, not confirmed.

Free runs two agents at once, a paid plan runs eight. The cost cap bounds work started, not turns already in flight — so an overrun of up to one turn per running agent is expected, and it is documented rather than pretended away.

Free, and where the ceiling is

	Free	Any paid plan
Chat	unlimited	unlimited
Agentic tasks	10 a day	unlimited
Tool-call rounds per task	10	50
Parallel agents in `/swarm`	2	8
Approval modes	all four	all four
API key required	none	none

The CLI is part of every paid plan rather than a separate subscription, and plans start at $4.99/mo. Paste the licence key from your plan into a session with /licence <key>.

The free caps are a JSON counter in your own home directory. Anyone who wants to delete it can — it is friction on a free tier, not DRM, and saying otherwise would be silly for an open-source CLI.

See plans → Read the docs →

Deploy from the same terminal

Write a Python function, deploy it to Animica Cloud, and get paid when people call it.

animica cloud login
animica cloud init            # writes a handler.py to start from
animica deploy handler.py     # validates, uploads, anchors on-chain, activates

Then animica cloud invoke, logs, status and earnings do what they say. animica deploy is a shortcut for animica cloud deploy.

Serving it yourself

The slowness above is a capacity problem, and the same package is the fix: animica up serves chat to the network and earns ANM for the work. A machine that can import torch and transformers starts advertising capacity on its own.

How mining works →