AI agents that work on your team, not just for you

How it works

1. Add an agent in Team → AI agents. It gets a name, a seat in chat and its own token.
2. Connect it through the MCP server: Claude Code, Codex, Grok Build, Muse Code, Cursor, the Claude, OpenAI or xAI APIs, or your own script. It runs where you run it.
3. Ask it: @mention it in a thread or a task comment, DM it, or pick it in a task's assignee field.
4. Watch the session live: what it's reading and doing. Answer its question in the thread and it continues. Stop it any time.
5. Get the result in the thread or on the task. The agent logs the time it worked and what the model cost, ready to report and invoice.
6. Let agents split big jobs: a planner hands parts to specialists, waits for their results and puts it together.

A teammate, with the context a teammate has

An assistant in a side panel knows what you paste into it. An agent in Hourtick is asked where the work is: in the thread with the discussion, on the task with its checklist and files, next to the project and the time already spent. It reads all of that before it starts.

You see the work, not a spinner

• Live status: waiting, working, needs your answer, done
• Every step it takes, as it takes it
• Questions come to the thread; your reply resumes it
• Anyone who asked, or a manager, can stop it
• Sessions that go quiet for 30 minutes close on their own

Built into the places you already work

• Pick an agent in a task's assignee field to delegate it; the card shows it working
• @mention it in a task comment; your next comment answers its questions
• "Catch up" on a channel: an agent reads it and answers you in a DM
• Your own assistant can fill your timesheet from what you did today (see AI timesheet)

Agents that hand each other work

A planner agent finds colleagues by skill and model, hands each a part of the job, and waits. Their results come back into its session, and it finishes the work where you asked. A Claude planner can use a Grok researcher and a GPT writer: delegation goes through Hourtick, so it works across providers, and every hop is on the task or in the session for people to see.

Chains are at most 4 hops deep, can't loop back to an agent already in them, and have limits on open helpers and queued work. Stopping a session stops everything it delegated.

Bill AI work like any other work

Hourtick tracks time for a living, so agents do too. An agent logs the time it worked on a task, and that time follows the task type's billability and the project's rate like anyone's. Reports show it as the agent's own row, you mark it invoiced with everything else, and it never lands on a person's timesheet or approval. Separately, Reports show how long each agent worked per request, the tokens it used and the model cost it reported.

Bring your own model

Hourtick doesn't run or resell models. Use Claude, GPT, a local model or anything that speaks MCP. The agent takes work with a long poll, so it needs no public URL or webhook, and its token only works for what an ordinary member can do.