Hands and permissions for AI agents on the Linux desktop.
The agents you already use work your apps and your browser through Cottus. Every action is checked against your policy, anything sensitive waits for your yes, and all of it is written down.
Works with Claude Code, Codex, Cursor and any MCP client. In early development, on GNOME first.
Short videos from a topic, made on your computer.
Our video studio. Claude writes the script and plans every shot, generative models make only the footage, and your computer adds the paper cut-outs, type, captions and the cut, in any of 34 edit styles. You see the cost before anything is paid for.
Launching soon. Early access by email.
Continuous security testing that runs inside your own network.
Our on-prem security product. Trinetra tests the systems you own on a schedule and at every deployment, points to the exact file and line behind each finding, and can draft the fix. Nothing about your code has to leave your network.
Coming soon. Early access by email.
Your agent asks.
Cottus checks.
You decide.
Here's how a session goes. Claude Code drives Text Editor by element id. Each call passes your policy, and the one that reads a document stops for your approval.
- youOpen Text Editor and start today's standup notes in a new tab.
- claudeapps_open{"app": "Text Editor"}allowed
- claudeapps_snapshot{"app": "Text Editor", "role": "button"}allowed
- [4.2] button "New Tab"
- claudeapps_press{"id": "4.2"}allowed
- claudeapps_snapshot{"app": "Text Editor"}allowed
- [5.7] text "" (0 characters, not shown)
- [5.3] text (1,204 characters, not shown)
- claudeapps_set_text{"id": "5.7", "text": "Standup, 26 Oct\n- …"}allowed
- youPull anything I left open from yesterday's notes.
- claudeapps_read_text{"id": "5.3"}asks you
-
claude-code wants to read the text of a document in Text Editor.
apps_read_text {"id": "5.3"} - approved by you, the agent reads it and carries on
- ~/.local/state/cottus/audit.jsonl
- {"ts_ms":1792991527412,"agent":"claude-code","capability":"apps_read_text","args":{"id":"5.3"},"outcome":"ok","detail":"approved by the user"}
Not another agent.
The layer under yours.
Cottus doesn't decide what to do. It gives the agents you already trust a way to reach your desktop, and you a say in every step.
Your agent connects
Claude Code, Codex, Cursor or any MCP client adds Cottus with one line, and sees every capability as a native tool.
claude mcp add cottus -- cottus mcp
Cottus checks every call
A small daemon on a per-user socket matches each call against your policy, asks you before anything sensitive, and appends it to the log under the agent's name.
It acts on your desktop
Apps through accessibility, by element id. Your own Chrome through an extension. Screenshots and real input only as a last resort, with the desktop's consent.
What agents can do today
- Desktop apps
- See a window's controls, press, fill fields and open apps, through accessibility. Ids like
3.12, not pixels. - Your browser
- Work in your own Chrome, with your logins, by element id. A site you haven't allowed asks first.
- Any other app
- A screenshot, then real clicks and keys, after the desktop's own consent dialog.
- Approvals
- Anything sensitive opens a dialog naming the agent, the action and its arguments.
- Watch mode
- Super+Escape stops every agent. They wait while you type, and the control about to be used is outlined.
- Dashboard
- One local page: what's waiting for you, what's running, what was held back, and what each agent did.
You write the rules.
Allow, ask or deny, for any capability, pattern, website or app. Reading and reversible changes are allowed by default; anything sensitive asks.
An exact rule beats a pattern. A broken file stops the daemon instead of guessing. Password managers and the keyring are always off limits.
[rules]
"apps_press" = "ask" # confirm app clicks
"web_click" = "ask"
"screen_*" = "deny" # no screenshots at all
"apps_snapshot" = "allow"
[sites]
allow = ["github.com", "linear.app"]
deny = ["mybank.example"] # never read or act here
[deny_apps]
names = ["Signal"] # never read or drive these
What Cottus will not do
- Read your documents in passing. A snapshot shows how long a document is, never what it says. Reading it takes your approval.
- Touch passwords or card numbers. Those fields are never read and never typed into. Cookies and browser storage aren't exposed at all.
- Overwrite your work. Writing goes into single-line fields and empty text areas. Replacing a document's text asks first.
- Let an agent approve itself. The dialog is drawn by your desktop. The dashboard can't approve a request or change the policy, on purpose.
- Listen on the network. The daemon uses a socket only you can reach, and the dashboard answers on 127.0.0.1 alone.
Where it is
Built one usable step at a time. GNOME on Wayland is the first target.
Working now
- 0Foundation. Daemon, policy, audit log, the
cottuscommand. - 1See and act. Snapshot an app's controls with ids, act by id, open apps.
- 2Agents connect. MCP server, approval dialogs for sensitive actions.
- 3Watch mode. Stop key, pause while you type, outline; screen and input fallback.
- 4Web. A Chrome extension in your own browser profile.
- 5Dashboard. Live activity, approvals, permissions, history.
Next
- 6Brain and memory. Plain-language tasks through your own agent's CLI; memory shared across agents; calendar, media and file search.
- 7Version 1.0. Packages, docs and a security review.
Questions
Is Cottus an agent?
No. It doesn't plan or decide anything. It's the layer the agents you already use go through to reach your desktop: it checks each call against your policy, asks you when it matters, and keeps the record.
Which agents work with it?
Any MCP client. Claude Code, Codex and Cursor connect with one command each, and every call is logged under the agent's own name.
How do I stop it?
Press Super+Escape or run cottus stop. Nothing acts again until cottus resume, and the stop survives a restart. Agents also wait while you're typing.
Can an agent approve its own request?
No. Approvals come from a dialog your desktop draws, which an agent can't answer. Deny, closing it, or no answer within 60 seconds all refuse the action.
What if an app has no accessibility support?
Cottus falls back to a screenshot and real pointer and keyboard input. All of it is sensitive, so it asks you, and the first use needs the desktop's own consent.
When can I use it?
It's in early development. Desktop apps, the browser, approvals and the dashboard work today. Follow @CottusAI for the first release.
More from Cottus AI
Open infrastructure for agents and models. Each project is public on GitHub under a plain license.
akernl
A self-hosted sandbox API. Run untrusted code in Firecracker microVMs with one HTTP call.
MIT · Python
Cottus Runtime
A C++/CUDA inference engine for Llama models, with Python bindings and its own KV cache.
Apache-2.0 · C++, CUDA
argus
Lightweight tracing for Python ML workloads. Spans in, a Chrome trace out.
Apache-2.0 · Python