Secure & collaborative agentic workspace.
Share an agent thread with your collaborators. You can write the next prompt together, comment on the agent's replies, and decide what the agent may do. Connections are peer-to-peer and end-to-end encrypted, so nobody else can read your threads. Models run locally through Ollama, so your work never has to leave the computers of the people in the thread. The agent's shell commands run with your own tools, inside your operating system's sandbox, and anything beyond the project's mode waits until the host allows it.
Cowork doesn't have releases yet, so you need to build it yourself. Install Rust and Ollama, then run:
git clone https://github.com/tidely/cowork
cd cowork
cargo run --releaseCowork connects peers directly over the iroh network. Collaborators join with the host's endpoint ID, a public key that identifies the host and secures the connection end to end. The host is in charge: everything goes through them, and they choose what each collaborator may do. Collaborators can read only, edit the draft, or, as admins, also send prompts, change the model, and approve the agent's tool calls.
The draft of the next prompt is a conflict-free replicated data type (CRDT), so everyone can edit it, comment on it, and attach files at the same time without conflicts.
The agent works on the host's computer with your own tools, so your shell, toolchains, and build systems work out of the box. Every command runs inside your operating system's own sandbox through mxc, so it can only change the project folders you share (and only in Write or Write + Network mode), can't touch your credentials, and has no network unless the project is in Write + Network mode. Commands that only read run without asking; anything that changes the project, or reaches the network, waits until the host allows it. The agent can also ask for more access than the mode, or for folders outside the project such as ~/.cargo, either for a single command or for the rest of its turn, which the host allows or denies.