How integrations work

Integrations connect your workspace — and your AI workforce — to outside services: Gmail, Slack, GitHub, Notion, Stripe, and roughly fifty more. They live in Settings → Integrations, and every card in that grid is a real, working integration. There is no mock catalog.

Every integration is a generated plugin

Cohort does not ship a fixed set of hand-written connectors. Each integration is a plugin that was authored by the platform's own build pipeline: it reads the service's API documentation, writes a manifest and sandboxed code, dry-runs every tool, and publishes the result to a registry shared across workspaces.

A plugin's manifest declares everything the platform needs to run it safely:

| Manifest part | What it declares | | --- | --- | | Tools | Typed, callable operations (e.g. send_message, list_repositories) with JSON-Schema inputs and outputs | | Auth | How the service takes credentials (apiKey, oauth2, basic, or none) and the exact fields to collect | | Settings | Non-secret options (text, number, boolean, select) that render as a form on the plugin's detail page | | Domains | The only hostnames the plugin's code is allowed to reach at runtime |

Three kinds of cards

The grid mixes three states, filterable with the chips above it:

• Installed — plugins this workspace has added. Open one to configure credentials, settings, and agent access, or to uninstall it.
• Available — plugins published to the global registry that this workspace has not installed. Open one to review its tools and add it.
• Connected — the state filter for anything currently wired up and working.

Category chips (Communication, CRM, Code & Repos, Finance, and so on) are derived from the plugins actually present — empty categories never appear.

Status, honestly

Each installed plugin carries a status pill derived from real state, never a decoration:

| Pill | Meaning | | --- | --- | | Connected | No credentials needed, or every required credential is sealed and stored | | Needs setup | The plugin requires credentials that have not been added yet | | Disabled | The install is switched off | | Removed | The plugin was yanked from the registry; existing installs keep working until uninstalled |

The security model in one pass

• Secrets are sealed. Credential values are encrypted with AES-256-GCM before they touch the database, and are never returned to the browser. See Credentials and agents.
• Code runs in a sandbox. Plugin code executes out-of-process in a separate VM realm with no filesystem, no environment, and no module loading. All network calls go through a guarded fetch that enforces HTTPS and the manifest's declared hostnames, with a 30-second wall clock and strict output caps.
• Agents need grants. Installing a plugin does not hand it to the workforce. You choose per plugin whether all agents may use it or only a named list. See Connect an integration.
• Every execution is audited. Each tool run appends a redacted event (plugin, tool, outcome, duration — never inputs, outputs, or secrets) to the workspace ledger.

Where the rest lives

• Procedures for installing, building, and configuring: Connect an integration
• Provider families: Google Workspace, Slack, GitHub
• Platform services with their own tabs and configuration: Agent email domains, Calls, Payments
• The API surface for your own systems: Build your own integration