TALOS
Docs 08

TALOS Code

An open coding agent that lives in your terminal. The thinking runs on the network; your files and shell never leave your machine.

Install

bash
npm i -g @talos-network/code

Requires Node 18+. Run it inside any project:

bash
talos-code                        # interactive, in your repo
talos-code "fix the failing test in test_api.py"   # one task, then exit

On first run it asks for an API key (get one at usetalos.xyz → settings → API keys) and saves it to ~/.config/talos-code/config.json. You can also pass it as the TALOS_API_KEY environment variable, or re-set it any time with /login.

How it works

Describe a task and the agent runs a loop with real tools: it finds the relevant code, reads it, makes a small targeted edit, runs your tests and stops once they pass. One command, no copy-paste.

  • It asks before every edit or command: allow once, allow for the session, or deny. Reads run automatically.
  • It shows a colored diff of every change.
  • Edits are small and targeted (line-range or search/replace, with a tolerant matcher), not whole-file rewrites.
  • Changes are syntax-checked and auto-reverted if they'd break the file.
  • It verifies its own work by running your tests. It won't call something done until it actually is.

Privacy

Your code stays where you are:

  • Edits and commands run locally. Your codebase is never uploaded to a sandbox. It's read and written on your own disk.
  • Secrets are stripped on your machine before anything is sent: API keys, .env values, private keys, emails. The model only ever sees a placeholder.
  • You can still ask the agent to wire a key into your project: the model sees the placeholder and the real value goes straight into your local .env, never crossing the network.
  • Nodes never learn who you are. A node is handed the snippet it's processing and nothing else: no account, no name.
Note
A node does see the snippet it's asked to process, same as any inference. What's different: no single company sees all of your code, ties it to you, or keeps a copy. The whole client is open source, so you can check exactly what leaves your machine.

Workspaces

Each project gets a .talos/ workspace. After every verified task the agent writes a short note to .talos/journal.md and reads it back on the next run, so it remembers past sessions and picks up where it left off. View it with /workspace. Commit it to share project history with your team, or gitignore it to keep it local.

It also reads project notes from talos.md, AGENTS.md, or CLAUDE.md in the repo root. Run /init to generate a talos.md for the current project.

Commands

CommandWhat it does
/initGenerate project memory (talos.md) for the current repo
/workspaceShow the project's work journal
/loginSet or replace your API key
/helpShow available commands
/exitQuit (Ctrl-C interrupts a running task; again at the prompt exits)

Configuration

All settings are environment variables:

VariableDefaultDescription
TALOS_API_KEYNoneYour TALOS API key (required)
TALOS_MODELtalos-heavyModel id: defaults to the unfiltered Heavy model
TALOS_YOLO0Set to 1 to auto-approve edits and commands
TALOS_API_URLhttps://usetalos.xyz/api/v1Override the API base
TALOS_MAX_STEPS40Max agent steps per task

Why it's different

Most coding agents belong to the company that built them. That company sees your prompts and code, sets your rate limit, can ban your account and can change the terms or shut down entirely, taking your workflow down with it.

TALOS Code flips that. The model runs across many independent GPUs run by people, so there's no single provider to throttle, ban, or take down. The models are unfiltered, so it won't refuse legitimate work. Your code stays on your machine and the agent itself is open source.

Previous ← Image generationNextAPI reference →