wobbli.es

The problem

Most AI coding tools wait for someone to prompt them. Teams still triage issues, chase CI failures, keep docs current and label PRs by hand. Wobblies watch those tools and do that routine work without being asked.

What I did

• •An event pipeline that verifies webhooks from GitHub, Slack, Linear, Sentry, Notion, Vercel and Jira, deduplicates them in Redis and routes them to agents by watch rules.
• •An activation worker running a native Claude tool-use loop, with models picked by tier, per-step budgets, deny rules and a full transcript of every step.
• •A sandbox verify loop: code-writing agents run setup, lint and tests in a devbox with up to three fix rounds, and only open a PR once tests pass.
• •A WOBBLIE.md format for defining agents in a repo, plus a library of 56 example agents for PR triage, flaky tests, docs drift and more.
• •A dashboard with activation timelines, approval gates from the web or Slack, dry-run mode, credit-based billing on Stripe and an MCP server.

How it works

A pnpm and Turborepo monorepo. The Next.js app runs on Vercel and a Hono API runs on Fly.io, with Postgres through Drizzle and Redis for BullMQ. Two queues split event routing from LLM execution, scheduled agents are triggered by cron, and finished activations go back into the router so one agent can trigger another.

Decisions and trade-offs

• •Replaced a single structured LLM response with a tool-use loop and step budgets. Denied tool calls go back to the model as structured refusals instead of being dropped.
• •Agent-written PRs are gated on passing tests in a sandbox, with automatic CI repair capped at two attempts per branch.
• •Credits are reserved up front and settled after each run by model and token use, and team membership is re-checked on every OAuth and API route.

Where it stands

In active development since June 2026, with close to a thousand commits. Free, Pro and Enterprise plans are live.