Open a project's dev environment and hit "Copy AI agent prompt" to try it. Changelog here: trigger.dev/changelog/ai-age…

The MCP server is useful after setup too. Ask your agent to filter runs, explain a failure from its trace, deploy, or query runs and LLM usage with TRQL, all from your editor.

It also carries the rules that keep generated tasks correct: export every task, use the built-in fetch, never wrap 𝚠𝚊𝚒𝚝.* or 𝚝𝚛𝚒𝚐𝚐𝚎𝚛𝙰𝚗𝚍𝚆𝚊𝚒𝚝 in 𝙿𝚛𝚘𝚖𝚒𝚜𝚎.𝚊𝚕𝚕. Browser login and your secret key still need you. The agent stops and asks.

No MCP? The prompt ships a full CLI path too. 𝚗𝚙𝚡 𝚝𝚛𝚒𝚐𝚐𝚎𝚛.𝚍𝚎𝚟@𝚕𝚊𝚝𝚎𝚜𝚝 𝚒𝚗𝚒𝚝, install the SDK, scaffold a task under src/trigger, start the dev server, confirm it registers. Same destination, more manual.

From there the whole loop lives in one conversation: ▪️ run 𝚝𝚛𝚒𝚐𝚐𝚎𝚛 𝚍𝚎𝚟 in the background, read the logs ▪️ inspect tasks and payload schemas ▪️ fire a test run, read the trace span by span AI spans carry model, token, and cost data.

With the Trigger MCP server connected, the agent calls 𝚒𝚗𝚒𝚝𝚒𝚊𝚕𝚒𝚣𝚎_𝚙𝚛𝚘𝚓𝚎𝚌𝚝. It detects an existing config, creates a project only if one's needed, and hands back the right setup steps for your dev environment.

New projects in @triggerdotdev now open with a "Copy AI agent prompt" button. Paste it into Claude Code or Cursor and your agent goes from an empty repo to a task that's registered and verified in the dashboard. How it works ↓

Setup a Smart column to point at a run's payload, metadata, or output with a JSON path: $.𝚘𝚛𝚍𝚎𝚛.𝚝𝚘𝚝𝚊𝚕 as a number $.𝚌𝚞𝚜𝚝𝚘𝚖𝚎𝚛𝙸𝚍 as text $.𝚜𝚝𝚊𝚝𝚞𝚜 as a badge Give it a label and it fills in down the whole list.

You can now show, hide, or reorder columns on the @triggerdotdev runs list. You can also add smart columns that pull a single value straight out of a run and display it for every row. ↓

Every environment in @triggerdotdev used to authenticate with a single secret key. Now each environment can hold as many keys as you want. Every service, CI job, and integration gets its own scoped credential. ↓

You can run bulk cancel and replay straight from the SDK. Point 𝚛𝚞𝚗𝚜.𝚋𝚞𝚕𝚔.𝚛𝚎𝚙𝚕𝚊𝚢() or 𝚛𝚞𝚗𝚜.𝚋𝚞𝚕𝚔.𝚌𝚊𝚗𝚌𝚎𝚕() at a filter and it hits every matching run, from a handful to millions. ↓

New in @triggerdotdev: the health report. One command tells you if your project is actually healthy: – is work starting? – are runs succeeding? – is your telemetry fresh? 𝚝𝚛𝚒𝚐𝚐𝚎𝚛 𝚛𝚎𝚙𝚘𝚛𝚝 𝚑𝚎𝚊𝚕𝚝𝚑 ↓

Define your prompts in code. Override them live when you need to. On @triggerdotdev a prompt is an id, a model, typed variables, and a template, and every deploy creates a new version. Override the text or model from the dashboard, no redeploy ↓

Reports: 𝚐𝚎𝚝_𝚛𝚎𝚙𝚘𝚛𝚝 returns a server-rendered health report for an environment. Is work flowing, are the runs that start healthy, is your telemetry fresh. In hosts that support MCP prompts, just run /𝚛𝚎𝚙𝚘𝚛𝚝 𝚑𝚎𝚊𝚕𝚝𝚑.

Agent chat: if you've shipped a 𝚌𝚑𝚊𝚝.𝚊𝚐𝚎𝚗𝚝(), your assistant can hold a real conversation with it. List the agents in your worker, start a stateful chat, send messages. Test an agent without building a UI first.

The MCP server for @triggerdotdev does a lot more than deploy and trigger tasks now. Your coding assistant can chat with your deployed agents, manage prompt versions, and pull a health report. All from Claude Code or Cursor. ↓

You could always see what a task did. Now you can see what it spent. Every LLM call your @triggerdotdev tasks make becomes a span in the trace. One flag, 𝚎𝚡𝚙𝚎𝚛𝚒𝚖𝚎𝚗𝚝𝚊𝚕_𝚝𝚎𝚕𝚎𝚖𝚎𝚝𝚛𝚢, on any Vercel AI SDK call. Nothing to install. Whole AI bill, or any single call, in the trace ↓

Every turn is a span in the dashboard: the prompts, the responses, the tool calls, how long each took. An AI metrics dashboard ships with every project too: spend, tokens, latency percentiles, cost by model. No instrumentation to write.

The turn is just a task. It takes the messages and returns a stream. 𝚜𝚝𝚛𝚎𝚊𝚖𝚃𝚎𝚡𝚝 on the server, 𝚞𝚜𝚎𝙲𝚑𝚊𝚝 on the client, and the API route that normally sits between them is gone. Only the new message goes over the wire. History lives on the server.

𝚌𝚑𝚊𝚝.𝚊𝚐𝚎𝚗𝚝 is a chat backend where every conversation runs on its own real machine. It boots on the first message, runs as long as the work takes, and keeps its memory between turns. Here's how one is put together, with @triggerdotdev ↓