Features

Built from the harness up

cobble doesn't wrap someone else's agent. Every layer, from the loop that runs a bot to who may open which computer, is ours, so it can be quick and strict at once.
The harness

Our own harness, not a wrapper

The loop that runs a bot is written from scratch. There is no borrowed agent runtime between your bot and its model, so every step carries less overhead.
Works with OpenRouter, OpenAI, Anthropic, Google, DeepSeek, Groq, Mistral, xAI, Ollama or any OpenAI-compatible endpoint
Falls back to the next model when one fails, and benches a provider whose key runs dry
Trims old tool output so long runs stay inside the model's context window
Stays on one model for a whole task unless that model fails
Models

A model for every job, per bot

A bot thinks with a text model, and calls on others to read images, transcribe voice, speak and make pictures. Pick each one per bot, or leave it on the defaults.
Five slots per bot: text, media reading, transcription, speech and image generation
Anything a bot leaves unset falls back to the install's defaults
A workspace's models win over the bot's own, for every run in that workspace
Limit a slot to one provider, keep fallback on, or turn a helper off
Sessions

Sessions, not one endless chat

A conversation with a bot is split into sessions, so its context never grows without bound. Each topic gets a clean start, and nothing earlier is lost.
Start a fresh session any time. It opens with a summary of the last one
A session that fills up rolls over on its own, summary and all
Only the open session goes to the model, so replies stay quick and on topic
The bot names every session. Search the old ones and reopen any to carry on
Every person has their own sessions with a bot, separately in each workspace
Computers

Every computer has an owner

A computer is a sandbox with a browser, terminal, code editor and desktop. A new bot gets no computer until you give it one.
Attach a computer to a workspace and every bot in it can use it
Or grant a computer to one bot, and to nothing else
Secrets are scoped the same way: everywhere, per workspace or per computer
Hidden secrets are never shown again, and are redacted from everything a bot reads
Memory

Memory in layers

A bot remembers what holds everywhere, what holds for you, and what holds in the workspace it's in. Each run loads only the layers that apply.
What a bot learns about you never reaches anyone else's conversations
Bots keep their memory up to date as they work
One SQLite database, backed up before every update and whenever you ask. Download any backup, or restore from one
Roles

Roles that contain the work

Super admins run the install, admins manage the crew, and members reach only what they're given.
Add a member to a workspace, or assign them one bot or a hand-picked few
Give those bots one computer, and that member's work stays on it
A bot's face is per person: "needs you" shows only to the one it needs
Read-only users can look around; revoked users can't sign in, and their history is kept
Features

And the rest of the crew's kit

What a bot needs to finish real work, and what you need to trust it with that work.

Take over a computer

When only you can do it, like a login or a captcha, a bot hands you its browser, terminal, code editor or desktop. Hand it back in chat.

A coordinator that delegates

It hands tasks to the right worker, which runs them in the background and reports back. With your say-so it adds bots and workspaces too.

Routines

Put a bot on a schedule and it queues its own brief on time. Set one up in chat or in settings.

MCP servers and custom tools

Add MCP servers, HTTP requests, scripts on a computer or prompts, for one bot, one workspace or the whole crew.

An activity log

Sign-ins, scheduled runs, cron and errors in one place, newest first, for the super admin.

Voice

Talk to a bot instead of typing, and have its replies read aloud.

On your phone

Install it as an app and get a push notification when a bot needs you.

Send it files

Drop in images, audio, PDFs, Word documents or CSVs. Documents are read on your machine, and a model that can't take images or audio gets them read first.

Meet your first bot

Add a bot, give it a job in plain language, and come back to finished work. On your machine, with your key.