Trigger.dev

The customer's agent code and its tools run on Trigger.dev rather than being supplied by it: a chat.agent calls streamText with tools the customer defines through the Vercel AI SDK, and any task can call an outside API through its SDK or fetch. No connector catalog or authenticated action framework of Trigger.dev's own is documented; the tools and credentials are the customer's code.

Sourcetrigger.dev/docs/guides/ai-agents/overviewread 2026-09-23

Trigger.dev is a workflow engine for long-running tasks: tasks trigger subtasks and batches and wait on them, with retries, queues, concurrency limits, idempotency, durable waits and versioned deploys, and its AI agent guides document prompt chaining, routing, parallelization, orchestrator-worker and evaluator-optimizer patterns, with chat agents running sub-agents.

Sourcetrigger.dev/docs/guides/ai-agents/overviewread 2026-09-23

No retrieval structure Trigger.dev maintains over the customer's corpus is documented. Its example projects build retrieval and chat over the customer's data in the customer's own code, and the chat transcript store holds conversations.

Sourcetrigger.dev/docs/guides/ai-agents/overviewread 2026-09-23

Agent actions can be held for approval in the AI Chat runtime: tools flagged as needing approval pause the agent's tool call until a person answers, the runtime carries that pause across idle gaps, redeploys and crashes, and its human-in-the-loop pattern gates new actions while a tool is waiting on the answer. The overview lists HITL approvals among what the runtime handles for the customer. The customer sets that hold per tool, and it applies to the agent's own action, not just a general durable wait.

Sourcetrigger.dev/docs/ai-chat/patterns/human-in-the-loopread 2026-09-23

The pricing page lists single sign on and role based access control on Enterprise, and Pro adds AWS PrivateLink. Enterprise also includes a SOC 2 report and a penetration test report, and a HIPAA BAA is available as an extra from the Hobby plan up.

Sourcetrigger.dev/pricingread 2026-09-23

Every run is recorded as a trace with logs and spans, and LLM observability turns each Vercel AI SDK call inside a task into a span carrying the model, provider, input, output, token counts, cost and latency, linked to prompt versions and queryable across runs. Chat sessions render in the dashboard's Sessions view, alerts go to up to 100 or more destinations, and log retention runs by plan (1 day on Free, 7 on Hobby, 30 on Pro, custom on Enterprise).

Sourcetrigger.dev/docs/ai/observabilityread 2026-09-23

A conversation outlives a single run, so the AI Chat runtime keeps each chat.agent conversation in a transcript storage keyed by chatId: when a run ends (a version upgrade, its turn limit, a crash), the next message boots a fresh run that reads the conversation back to answer in context. The runtime writes the transcript from the agent's own turns, including tool calls and results. It is keyed per conversation.

Sessions carry an expiresAt hard retention deadline, and streams age out on the plan's retention window. It is read back through the loadTranscript and chat.history APIs and the dashboard's Sessions view, and changed through actions (undo, edit, regenerate) or a TranscriptStorage adapter to the customer's own database.

Sourcetrigger.dev/docs/ai-chat/transcript-storageread 2026-09-23

Trigger.dev is open source under Apache 2.0 (per the repository's license file) and self-hosts in the customer's own infrastructure under a documented self-hosting guide, alongside the managed cloud, where runs can state a region preference.

Sourcetrigger.dev/docs/self-hosting/overviewread 2026-09-23

Ready made starting points ship with the product: agent pattern guides (routing, translate and refine, content checks), dozens of example projects (a chat agent over ClickHouse, a Cursor background agent, a batch LLM evaluator, human-in-the-loop workflows, Mastra agents with memory), and building-with-AI resources for coding agents. These are examples a developer builds from, not a catalog of agents a buyer adopts.

Sourcetrigger.dev/docs/guides/ai-agents/overviewread 2026-09-23

Work wakes on its own: scheduled tasks run on cron schedules declared in code or created from the dashboard or API (10 on Free up to 1,000 or more on Pro), and a chat agent wakes when a new message arrives on its session.

Sourcetrigger.dev/docs/tasks/scheduledread 2026-09-23

The customer chooses the model through a surface of Trigger.dev's own. AI Prompts define prompt templates with a default model in code, version them on each deploy, and let the team create overrides from the dashboard that change the prompt text or the model without redeploying, with token usage, cost and latency tracked per prompt version; chat agents resolve the named model through a provider registry.

Sourcetrigger.dev/docs/ai/promptsread 2026-09-23

Trigger.dev is callable from outside through its TypeScript SDK and management API for triggering and inspecting runs, the CLI, and an official MCP server whose tools let coding assistants list, start and message chat agents and manage runs.

Sourcetrigger.dev/docs/ai-chat/mcpread 2026-09-23

Testing here covers the agent's code, not evaluation of its output. The AI Chat testing harness (mockChatAgent) runs the agent's real turn loop, lifecycle hooks and actions in memory with the model mocked, so a test asserts on chunks whose content the test itself scripted. That checks the plumbing around the model rather than what a real model produced. AI Prompts track token usage, cost and latency per prompt version, which compares spend and speed rather than quality.

Sourcetrigger.dev/docs/ai-chat/testingread 2026-09-25

Browser workloads run in an environment Trigger.dev hosts: tasks can install system packages and headless browsers through build extensions, and an example project drives a hosted browser from Anchor Browser inside a task. The control of the browser comes from those libraries and services in the customer's own code, and no Trigger.dev API for screenshots, clicks or typing is documented.

Sourcetrigger.dev/docs/guides/example-projects/anchor-browser-web-scraperread 2026-09-23