nanobots

Self-improving agent loops for any GitHub repo. An outer loop triages, prioritizes, dispatches, reviews, and learns; nanobot workers claim cards and ship PRs. GitHub is the only state store — issues are the intake, a Projects v2 board is the kanban, PRs are the audit trail, and a pinned issue is the heartbeat. When the bots are out of their depth, they summon-human.

Inspired by Autoresearch/Introspection: the durable product isn't the code, it's the loop that maintains it — with humans as a deliberate signal source, not a bottleneck.

                 ┌────────────────────────────────────────────────┐
   signals ───▶  │  OUTER LOOP (control + learning)               │
   issues,       │  ingest → triage → prioritize → dispatch       │
   errors,       │  → review outcomes → LEARN (update its policy) │
   humans        └───────────────┬────────────────▲───────────────┘
                                 │ dispatch        │ outcomes (PRs, CI)
                                 ▼                 │
                 ┌────────────────────────────────┴───────────────┐
                 │  NANOBOT WORKERS (inner loop)                  │
                 │  claim card → branch → build → tests →         │
                 │  PR (Closes #N) → gates → merged by outer loop │
                 └────────────────────────────────────────────────┘

Install (any repo)

cd your-repo
npx nanobots-sh init        # or: curl -fsSL nanobots.sh/install | sh

init asks ~6 questions (test commands, hard-gate areas, WIP cap — with answers auto-detected from your repo where possible; no model/API key needed to install), then:

1. renders the loop into .nanobots/ (prompts, triage rubric, recipes, learnings log, runner script, config.json),
2. writes GitHub workflows + issue intake forms,
3. scaffolds GitHub state via gh: the project board, Priority/Size fields, labels, and a pinned status issue,
4. prints the two manual steps GitHub's API can't do, plus the secrets checklist.

After init the repo is fully self-contained — markdown prompts, one shell script, plain workflows. No runtime dependency on this package. Your repo's rubric and recipes are supposed to drift from the templates: that's the learning.

Optional: init --smart additionally uses the claude CLI (if installed) to study your repo and draft repo-specific recipes and gates into .nanobots/SUGGESTIONS.md.

Run it

The loop is stateless: "running" = something executes one cycle on a cadence. Run either role anywhere — see .nanobots/RUNTIMES.md after install.

Where	How
Laptop (interactive)	/loop in Claude Code with .nanobots/LOOP-PROMPT.md (outer) or WORKER-PROMPT.md
VM / your own box	npx nanobots run outer (or bash .nanobots/run-cycle.sh outer) on a systemd/launchd timer
GitHub Actions	the installed nanobots-outer.yml cron (set NANOBOTS_OUTER_ENABLED=1)
Sandbox cloud (Daytona/E2B/…)	same script + two env vars; a ~20-line launcher when you need per-item isolation

Billing is an env var, not a feature. CLAUDE_CODE_OAUTH_TOKEN runs everything on a Claude Pro/Max subscription (mint once: claude setup-token). ANTHROPIC_API_KEY is metered. Any Anthropic-compatible provider works — e.g. DeepSeek via ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic. Mix per runtime: cheap model workers on a VM, frontier model for the outer loop's judgment.

Workers are swappable at the PR seam. A worker is anything that reads an issue's work-spec and opens a PR with Closes #N — Claude Code (ours), GitHub Copilot's coding agent (assign the issue), Codex, OpenHands. The board can't tell the difference.

How the learning works

Every finished (or failed) work item gets an entry in .nanobots/LEARNINGS.md — an append-only memory. Every ~10 entries the outer loop runs a distill pass, promoting durable lessons into the triage rubric, the recipes, and your agent instructions file. The loop's policy documents are its weights; GitHub history is the training log.

Requirements

• gh CLI authenticated, with the project scope (gh auth refresh -s project)
• The Claude GitHub App on the repo (claude /install-github-app) — only needed for the Actions execution paths
• Two secrets for Actions mode: CLAUDE_CODE_OAUTH_TOKEN, and PROJECTS_PAT — a classic PAT (project + repo) from a human account. Two hard-won facts baked into the design: the default GITHUB_TOKEN cannot touch org Projects v2 at all, and claude-code-action refuses runs initiated by bot actors, so dispatch comments must post as a human.

Commands

npx nanobots-sh init [--smart] [--no-github] [--yes]   # scaffold a repo
npx nanobots-sh update                                  # re-render engine-owned files only
npx nanobots-sh run <outer|worker>                      # one headless cycle

update never touches repo-owned files (TRIAGE.md, RECIPES.md, LEARNINGS.md, config.json) — those belong to your repo's loop. Engine-owned files carry a nanobots:engine-owned marker in their first lines.

Browser extension (signal capture)

extension/ is a zero-dependency MV3 Chrome extension that feeds the loop from any webpage: click → screenshot → annotate (pen/box/arrow) → filed as a nanobots:inbox issue the board auto-adds for triage. Screenshots live in your Cloudflare R2 bucket (free tier; the issue embeds the link, git stays binary-free) — capture is disabled until R2 is connected, and the options page walks through the 3-minute free setup. It keeps a local history of everything you've filed (with live state), and has a repo chat — a BYO-model agent with real repo tools (code/issue search, file read, report filing with your pasted screenshots). The chat agent's prompt lives in the repo (.nanobots/EXTENSION-PROMPT.md) and is refined by the outer loop based on how extension-filed reports fare in triage — the intake itself self-improves. Install: chrome://extensions → Developer mode → Load unpacked → extension/. See extension/README.md.

Docs

• docs/architecture.md — the two loops, the GitHub surface mapping, the cycle spec, design decisions
• docs/research/ — the July 2026 tooling landscape + the GitHub Projects v2 automation reference this design is built on

License

MIT