What is Nimbus?

Nimbus is an autonomous software engineering agent. Give it a task in plain English — it indexes your codebase, generates a precise implementation plan, writes the code, runs your test suite, and opens a pull request.

It runs locally on your machine or on the hosted backend at api.get-nimbus.com.

The pipeline

Every Nimbus task follows the same phases:

1. Clone — isolate the workspace (hosted) or detect the repo (local)
2. Index — embed all source files using voyage-code-2 + BM25 into ChromaDB
3. Plan — Claude Opus retrieves context and generates a file-level change plan
4. Approve — plan shown with confidence score; user approves before execution
5. Implement — Claude Sonnet executes the plan via agentic tool-use loop
6. Verify — run the actual test suite (pytest, tsc, eslint, cargo)
7. Fix — if verification fails, Nimbus regenerates the plan from error output (up to 5 iterations)
8. Diff preview — full diff shown before anything is pushed
9. Review — Claude Sonnet self-reviews the diff and posts a structured critique
10. PR — branch pushed, pull request opened, comments monitored
11. Cleanup — workspace cleaned up
12. Memory — task outcome written to per-repo memory for future tasks

Two modes

Local mode — runs on your machine. Indexes your files, writes changes directly to disk, commits locally. Launched with nimbus.

Hosted mode — runs on Railway. Your GitHub repo is cloned, the task executes, and a PR is opened. Launched with:

nimbus run "add rate limiting to the API" --backend https://api.get-nimbus.com

Philosophy

The knowledge required to write and maintain production software is enormous. Most of it is mechanical. Nimbus automates the mechanical parts — retrieval, implementation, verification, review — and leaves the decisions to you.

Every task requires your explicit approval at two gates: the plan and the diff. Nothing is pushed without your sign-off.