AI Agents (MCP)

Taskforce.sh lets you add AI agents as first-class, non-human members of your organization. An agent connects to your queues over the Model Context Protocol (MCP) and can investigate and act on them, list failed jobs, retry the ones that are safe, pause a queue, and more, bounded by the exact same permission model you use for human teammates, and with every action recorded in an audit trail.

You can connect any MCP-capable client, such as Claude Desktop, Claude Code, Cursor, VS Code, or your own agent built with an MCP SDK.

Why agents?

Queues fail in ways that are tedious for humans and well-suited to automation: an SMTP outage fails a batch of welcome-emails, a downstream API blips and a handful of jobs need retrying, a queue backs up overnight. An agent can triage these the moment they happen, and, crucially, it does so within guardrails:

• It only sees the queues on the teams it belongs to.
• It can only perform the actions its role grants (view, retry, delete, …).
• Everything it does is logged, with an allowed / denied outcome.

How it works

1. Create an agent in your organization and add it to one or more teams, just like inviting a person. The agent inherits that team's role (and any custom-role overrides).
2. Mint a token for the agent. The token is shown once; you paste it into your MCP client as a Bearer credential. Tokens can be read-only or full, and can be given an expiry.
3. Point your MCP client at the Taskforce.sh MCP endpoint. The agent can now call tools scoped to its permissions.
4. Review the activity log for the agent at any time, and export it if you need a record.

   MCP client                Taskforce MCP            Your queues
 (Claude · Cursor · ──▶  ── endpoint (/mcp) ──▶  permission check ──▶  action
  custom agent)           Bearer tfa_…             per team & queue        │
                                                        │                  ▼
                                                     denied ──────────▶ audit log
                                                                           ▲
                                                        allowed ───────────┘

Next steps

• Managing agents — create agents and assign team access.
• Tokens — mint read-only or full tokens with optional expiry.
• Connecting an MCP client — Claude, Cursor, or your own.
• Permissions — what an agent can and cannot do.
• Tools reference — every tool the agent can call.
• Activity & audit log — review and export what an agent did.