trigger event → sandboxed agent run → code
Serverless agent graphs for coding.
A PR, Linear, or Slack event fires the trigger. Each run gets an isolated sandbox that executes your graph of custom agents end-to-end and ships the output back where work started. Every step streamed live, fully queryable after.
Build your own Greptile
Full-context code review on every pull request.
Each of these is just a few nodes: a trigger, agents, version control. The difference is the agents are yours. Their instructions, skills, and tools are specialized to your codebase and your conventions, not a one-size-fits-all product you can't open up.
You are the team's merge doctor. Rebase onto the base branch, resolve conflicts preferring semantic intent over textual order, and never touch generated files. If both sides changed behavior, stop and describe the collision instead of guessing…
Agents that know
your codebase.
A custom agent is standing instructions, skills, and tools packaged under a name. Skills teach it your conventions, from commit style to how your modules are laid out. Tools define exactly what it's allowed to do: built-in actions like posting a PR review or replying in Slack, plus commands you author yourself.
Each agent also picks its runtime: which harness executes it and which model it thinks with, metered through the platform or running on your own Claude subscription. Define it once and drop it into any workline, as a pipeline step or live in a Slack session. Its knowledge, permissions, and runtime travel with it.