Someone types /request new in Discord and describes what they want. The bot opens a GitHub issue and posts an Approve/Reject card. On approval, Claude Code builds it in GitHub Actions and opens a pull request into the group’s own module repository, and the bot posts the link back and @-mentions whoever asked.
Most “AI builds your feature” demos skip the three parts that make it safe to actually run, and this one treats them as the product:
- A human approval gate. Only someone granted
builder.approvecan start a build. Each approval is signed, single-use and expires after 15 minutes. - Containment of untrusted input. Request text is redacted, rate-limited and fenced off as data inside the agent’s prompt, so it can’t override the repository’s rules.
- An answer to who pays. Each requester stores their own Anthropic credential once, sealed into a per-user GitHub environment the bot can’t read back. There’s no owner bypass, and approval checks the day’s build count before anything runs.
One instance, one trust domain
Guildwright is an engine plus a template. A group makes one private module repository from the template, registers its own GitHub App and runs its own bot from that repository. Modules are in-process code, so merging one gives it the bot’s full authority. That’s why there’s no hosted bot that runs other people’s modules, and deliberately no way to build one.
Requests can be listed, cancelled (which also stops the GitHub run, so it stops spending credits) and retried. guildwright doctor reports every configuration problem at once instead of one per restart.