Using Cohent | Cohent docs

Cohent coordinates work started independently by different developers. Keep using your own coding agent. You do not create a Cohent task, assemble a team of agents, or copy conversations between them.

Connect once

Sign in to Cohent and choose your repository in Account settings.
Open Connect agent and choose the integration you use.
Open Cohent Connect on your Mac. Choose your local repository folder and confirm your team's Cohent address.
Review what the integration reads, shares, changes and can block. Explicitly approve, then choose Connect in browser. Approve this repository in the browser.
Back in the companion, choose Review exact changes, then Install integration.
Open your coding agent normally and complete its own hook trust review.
Choose Check connection & coverage in the companion.

The companion bundles its runtime. It does not launch an agent or approve host trust. It uses its bundled adapter, not executable setup code from your repository. Existing unrelated agent settings are preserved.

The current macOS companion is a development build. A signed/notarized public download is not yet configured. The web panel says so rather than offering a dead download. Windows/Linux desktop installers are not implemented. See desktop setup for the build and release boundary.

What happens during work

An enabled hook announces a small, locally minimized initial intent. Your agent maintains more specific structured facts as its plan becomes clear: outputs, contracts, assumptions, components and dependencies. An initial short goal alone may not reveal enough to find a relationship.

Cohent compares newly revised work against indexed, live candidates in the same project. It does not connect every pair of agents. A billing provider and a checkout consumer can become related because they name the same contract even when their files are different.

A useful relationship carries its reason, evidence, type and severity. Only relevant facts are routed to its participants. Unrelated agents receive none of those packets. Peer text is always untrusted collaboration data, never a system instruction or permission grant.

Participants can acknowledge relevance, request context, agree on a contract, report a resolved blocker, or request human judgment. A changed plan causes a fresh evaluation; an agreement for an old version cannot approve the new one. Resolved relationships leave concise receipts rather than permanent group chats.

The Control Room

Overview: current work grouped by person, meaningful live relationships, decisions needing attention, integration degradation and recent outcomes.
Work: what each connected developer's agents are attempting.
Coordination: evidence, minimum shared context, participant responses and receipts. Technical JSON is secondary to readable facts.
People: actual members and observed sessions, not inferred hidden activity.
Notifications: requested human decisions, guarded write blocks and serious connection/trust problems. Routine exchanges remain in activity.
Project: connection, policy and access settings.

When a relationship explicitly needs a human, its participant can record a decision in the GUI for their own work. This does not impersonate the other developer or approve their side of a contract.

Mark a relationship useful or noisy to help evaluation. Did Cohent miss a relationship? records two observed intents and a reason; it does not force a connection or authorize additional sharing.

What protection means

Installation, trust, liveness and actual tool coverage are different things.

Claude Code and Codex can enforce declared scope for supported, visible edit targets when the relevant trusted hook actually runs. Unknown edit payloads are denied at that hook boundary. Pause/cancel and unavailable authority are also checked at announced boundaries.

Cohent does not inspect arbitrary shell effects or promise that background processes, unannounced tools or every keystroke are guarded. Codex hook changes can invalidate trust. Claude trust cannot be verified outside an observed hook. The diagnostics explicitly report unknown or unavailable capabilities.

Cursor and Copilot are advisory participants; the desktop installer does not configure them or promise write blocking. Agent-side adapter operations remain available under the host's explicit permissions.

A semantic relationship is normally a reason to exchange context, not a blanket write prohibition. Concrete enforceable path conflicts remain a guard invariant.

Disconnects and failures

A session needs a live lease to look controllable. Sleeping or crashed sessions become idle/stale. Losing project access stops further coordination access. New device-bound credentials let one device disconnect without revoking another current device. Old credentials without attributable device ownership are revoked conservatively for that user and repository.

If the service is unreachable, guarded actions that require live authority fail closed. Advisory integrations cannot promise that writes are blocked.

Remove an integration

In Cohent Connect choose Remove integration. Confirm removal of the selected agent's Cohent-owned hooks. Other handlers and settings remain.

Choose Disconnect this device separately to revoke its repository access. Pinned executable caches remain recoverable locally. Neither action grants host trust, deletes your repository, or removes another developer.

Advanced fallback

For repositories already containing the adapter, the Connect agent panel includes a collapsed command-line fallback. Agents can use the same protocol internally; developers do not need to manage it for normal coordination.

Cohent does not execute repository jobs, replace Git/CI/code review, or manage a swarm. Historical task simulator records remain read-only and explicitly labelled.