Opencastle

OpenCastle

  What your team's AI learns stays with the team — in every teammate's assistant, reviewed like code.

  Website ·
  Quickstart ·
  Docs ·
  Architecture ·
  Contributing

Every AI coding assistant now remembers what it learns: the build that needs a
flag, the convention nobody wrote down, the staging API that rate-limits. And
every one keeps it to one person — Claude Code and Codex on a laptop, Devin and
Cursor on their servers. On one of our projects, Claude Code held 87 memories
about the code in October 2026, and not one had reached a teammate. A new
teammate's assistant starts from nothing, and the next person to hit the same
problem pays for it again.

OpenCastle makes it the team's. Instructions, skills and agents live in the
repository and compile for the assistant your company uses. What agents remember
about the project becomes a lesson when the agent stops (a hook in Claude Code
and VS Code), reviewed in the pull request with the work, and every teammate's
assistant loads it before its next task. What they remember about you stays
yours. Across repositories, the organisation's standard is an
Agent Plugin every repository extends.

It starts from your code: the stack with its versions, the commands, the routes
and the models, so the assistant knows the project on its first task.

With the tools you have, or on its own. OpenCastle runs no agent, no
infrastructure and no search; it is the team's layer under the ones you already
pay for. It reads what Claude Code, VS Code and Codex remember on your machine,
and writes into each assistant's own format. Devin installs your baseline as a
plugin, and platforms such as Linear, Sentry or Notion reach every assistant as
integrations. On its own it needs one repository, git and…