Coworker Agents overview
Run autonomous agents that work alongside your team — answering Slack threads, handling scheduled chores, and opening pull requests — while MintMCP manages their credentials, tool access, and audit trail.
Coworker Agents is in beta and available on Enterprise plans. Contact us to enable it for your organization.
How it works
A coworker agent is a directory in a GitHub repository you connect to MintMCP. Everything the agent is — its configuration, instructions, and memory — lives in that directory as files you can read, review, and edit like any other code. There's no hidden state: the repository is the source of truth.
| File | Purpose |
|---|---|
agent.yml | Configuration: triggers, model, tools, secrets, sandbox settings |
CLAUDE.md | Instructions — the agent's role, goals, and rules |
progress.md | Persistent memory the agent updates across runs |
activity-log.csv | Append-only audit trail the agent maintains itself |
inbound/ | Drop folder where people (or other agents) leave tasks |
Each run starts a fresh session on MintMCP's hosted runner: the agent reads progress.md and inbound/, does its work, updates its memory and activity log, and commits the results — directly for changes inside its own directory, or as a pull request for anything else. Because memory lives in the repository, the agent picks up where it left off on the next run.
Runs execute in an isolated sandbox with restricted network egress, and compute disappears when the run ends. See Coworker Agents security.
Triggers
An agent can have any combination of three triggers:
| Trigger | How it fires |
|---|---|
| Slack mention | @-mention the agent's Slack app; it reads the thread and replies in it |
| Schedule | Cron expressions (UTC) for recurring work |
| Manual | The Run button in MintMCP, with a prompt and model choice |
What agents can do
- Converse in Slack — each agent gets its own Slack identity, acknowledges mentions immediately, and posts progress in the thread. See Connect a Coworker Agent to Slack.
- Call your tools through MintMCP — connect a vMCP to give the agent scoped, audited access to your MCP servers. Tool access is a real allowlist: a tool not in the agent's
allowed_toolsis not callable, even if its server is configured. - Ship changes as pull requests — agents work on branches and open PRs for review, and leaving PR feedback becomes input for the agent's next run.
- Delegate to sub-agents and reuse skills — agents can define sub-agents for focused work and share skills across your organization.
Governance
The governance model is the same one the rest of MintMCP uses, so agents stay visible and controllable:
- IT & Security teams get an audit trail on every gateway tool call, session visibility through Agent Monitor, and rules that can flag or block risky actions.
- Admins control which repositories agents live in, which tools each agent can call, and how its secrets are stored and brokered.
- Anyone with repository access can read an agent's instructions, memory, and full run history — an agent is reviewable code, not a black box.