Email content

Trigger.dev update – August 2026

Hey there, 👋

Here's what's new in Trigger.dev this month. Check out our full changelog for more details.

Introducing chat.agent

Build durable AI chat that runs on a real Linux machine with no turn timeouts, keeps streaming through refreshes and crashes, and sleeps between messages so waiting costs nothing.

Every conversation gets its own machine that holds your variables, caches, and sub-agents in memory across turns, then wakes up exactly where it left off, even days later. Write the turn as a Trigger.dev task and point the AI SDK's useChat straight at it, with no API route in between. Learn more

Multiple API keys per environment

Create as many keys as you need per environment, each scoped to a specific slice of the API (trigger only, deploy only, observer, and more) and optionally locked to as few as one task, up to 10 per key. Give a key an expiration date or revoke it instantly, and because keys are independent, retiring one leaves every other key untouched with no environment-wide rotation. Learn more

Bulk actions from the SDK

Cancel or replay runs matching a filter straight from your backend with runs.bulk.replay() and runs.bulk.cancel(), from a handful to millions at once. The filter is the same shape as runs.list(), so you can select by status, task, tags, or version, then poll the returned action for its status and counts. Learn more

AI prompts: code-defined, versioned, overridable

Define prompts as typed exports in your code, with a model, Zod-typed variables, and a Mustache template, so a missing or wrong-typed variable is a compile error instead of a broken generation you find in production. Every deploy snapshots them, and you can change the wording or swap the model from the dashboard without redeploying, with every generation tracing back to the exact version that produced it. Learn more

The health report: diagnose a project in one command

Get a server-computed verdict on whether work is starting, whether the runs that start are succeeding, and whether your telemetry is fresh, each with its own headline and a suggested next action. There's no model in the diagnosis, so the same inputs always produce the same answer, and you get it from the CLI, the get_report MCP tool, or the API. Learn more

MCP server gets agent chat, prompt management, and reports

Your coding assistant can now hold a real conversation with any chat.agent you've deployed, manage prompt versions end to end, and pull a health report for an environment. It also queries your run data with TRQL, our SQL-style query language, and drives the dev server, all running at @latest with nothing to upgrade. Learn more