MCP server for AI agents
Run your agents from the assistant you already use.
Connect Claude Code, Claude Desktop, Cursor or any MCP client to your workspace. Ask which agents failed and why, decide approvals, invoke agents and set schedules, with the same permissions you have in the dashboard.
The problem
The questions come up where you are working, not in the dashboard.
You are in your editor or in a chat with Claude when you wonder whether last night's runs worked, or a teammate asks why the outreach agent went quiet. Answering means switching to another tab, finding the agent, opening the runs and reading the events. It is small, and it happens many times a day.
- Context switching. The answer sits in the dashboard, and the question comes up somewhere else.
- Small changes, many clicks. Pausing an agent, moving a schedule or rerunning a test case is quick to ask for and slow to do by hand.
- Scripts per task. Automating a check means writing and keeping a script against an API.
How it works
One URL, one key, every tool.
AgentOS runs a remote MCP server. Any client that speaks the Model Context Protocol over Streamable HTTP can connect to it: Claude Code, Claude Desktop, Cursor and others.
- Generate a key. A personal key acts as you, with your own role, in every workspace you belong to. A workspace key acts as an admin of one workspace, for shared automation.
- Add the server to your client. One command in Claude Code, or a few lines of JSON in Cursor or Claude Desktop.
- Ask in plain words. The assistant picks the tools: fleet health, runs and their events, approvals, schedules, evals, reports, alerts and more.
- Check what it did. Every run it starts is recorded like any other, and key, member and agent changes land in the workspace audit log with the key that made them.
Ask
Which agents failed this week, and why?
The assistant reads each agent's health over the days you ask about, worst first, with its failure rate and the trend. Then it opens the failed runs and reads their events, down to the tool call that returned the error.
The answer comes back in the conversation, and the record it came from is the same one you would open in the dashboard.
Act
Decide, invoke and schedule without leaving the conversation.
- Decide approvals. List what is waiting for a person, read the exact call the agent wants to make, then approve or reject it with a note.
- Invoke hosted agents. Run an agent on the input you give it and read its output, with conversation memory when you want follow-ups.
- Set schedules. Put an agent on a daily or weekly schedule in your time zone, change it or remove it.
- Test before you ship. Create eval cases, run them and read which assertions failed.
- Build agents. Create an agent, set its prompt, model and tools, file it in a folder, or install one from the marketplace.
- Get a report. Ask for a summary of one run, or a period report on an agent.
Same rules
Your assistant has your permissions. No more.
A personal key carries your live role. A member reaches only the agents they created or were granted, and workspace settings stay with owners and admins. Remove someone from the workspace and their key stops working there on the next call.
Connecting an assistant does not bypass approvals. A tool that needs a person's sign-off still pauses the run and waits, however the run was started. Agent keys cannot reach the server at all.
For developers
The same tools, over plain HTTP.
Every tool except creating a workspace is also a JSON endpoint, with the same keys, schemas and role checks, and the TypeScript and Python SDKs wrap them. Use MCP from your assistant, and the control plane from scripts and services.
Go deeper
How it works, in detail
More in Run
Put agents to work on a schedule, with your tools and your documents.
Give one job to an agent this week.
Start with one repetitive workflow, gate the sensitive step behind your approval, and read the first run end to end.