Skip to content

About

⚔️ Command & Control center for a pool of VPSes + Claude subscriptions: coding at super scale. 1 VPS = 1 team; dispatch /goal missions, merge PRs sequentially with claudetm, verify deploys.

Topics

Resources

Stars

3 stars

Watchers

0 watching

Forks

Repository files navigation

Claude and Conquer

Claude and Conquer

Command & Control center for a pool of VPSes + Claude subscriptions — coding at super scale.

One commander (you + Claude on the main box) directs a fleet of worker VPSes. The main box runs this repo and mirrors the whole workspace at ~/workspace/<org>/<repo>, so the commander can read the real code locally. Each worker is one team: a coding agent ready (claude = Anthropic Max, or glm = z.ai GLM), repos cloned, claudetm installed. You paste a /goal, the commander picks the idlest box, makes sure the code + secrets are there, dispatches the mission, PRs merge sequentially, and the flight controller verifies the work actually deployed.

        ┌─────────────────────────┐
        │  COMMANDER (main box)   │  /goal … → cnc goal   (runs this repo)
        └────────────┬────────────┘
      ┌──────────────┼──────────────┐
      ▼              ▼              ▼
  ┌────────┐     ┌────────┐     ┌────────┐
  │worker-1│     │worker-2│     │worker-3│   1 VPS = 1 team = 1 coding-agent sub
  └───┬────┘     └───┬────┘     └───┬────┘
      ▼              ▼              ▼
  parallel agents → PRs → claudetm merge-pr (sequential) → configured deploy → deploy-check

The idea

  • 1 VPS = 1 team, running a coding agent — claude (Anthropic Max) or glm (z.ai GLM).
  • Every box mirrors the whole workspace at ~/workspace/<org>/<repo> — identical to local.
  • C&C makes sure the stuff is there: every repo cloned + its .env delivered from Bitwarden (workers can't reach the vault — only the main box can). The agent does the rest itself: deps, build, tests, PRs.
  • The commander picks the box: cnc goal round-robins by state/load across ready boxes — any box can run any goal, since all mirror the workspace. --team / --pool override.

Workflow

Inside a Claude session on the main box, use /goal <text> — the commander enriches the goal with standing orders (parallel agents, sequential claudetm merge-pr, verify gate, ship to the project's configured deployment method, 100% completion) and dispatches it to the idlest ready box. That box's agent does the deep, full work. The main agent just launches the agents on the worker VPSes.

Projects configure one of the supported deployment methods: gitops-argocd, docker-compose, or coolify. GitOps projects may use an infrastructure repository and stack; other methods use their own configured target.

/goal use parallel agents, claudetm merge-pr each PR, deploy using the configured project method —
      finish the dashboards, find bugs, refactor, 100% complete

How the commander picks a box. Since every box mirrors the whole workspace, cnc goal round-robins by state/load — it picks the least-busy, coolest-burn ready box and pre-flights that box's active 5h window via ccusage, refusing a hot subscription (raise CNC_BURN_LIMIT or --force to override). Right before handoff it makes sure the stuff is there (clone if missing + deliver the .env), so the agent just codes. Because the main box also holds the ~/workspace clones, the commander can read the real code locally to consult while missions run elsewhere. Each box can run different coding agents (cclaude = Anthropic Max, cclaudez = z.ai GLM, …) — see docs/agents.md.

Fork it

Fork this repository, clone your fork, start Claude, and run /setup — Claude interviews you for servers + projects + orgs, writes the inventory, provisions every VPS, and hands you the auth checklist. The only step that can't be automated is the Claude subscription login:

bin/cnc ssh <team> --login            # /login the Claude subscription on that box

GitHub access is propagated for you: bin/cnc gh-setup <team> copies the main box's ssh key + gh token so each worker clones and pushes with the same access as the commander (cnc provision runs it automatically).

Prerequisites

  • An operator machine with authenticated gh and Bun installed.
  • SSH-reachable Ubuntu or Debian worker VPSes with passwordless sudo.
  • One Claude Max login for each worker box; a human completes each login with bin/cnc ssh <team> --login.

Security

bin/cnc gh-setup copies the operator's SSH key and gh token to each worker so it can access the same repositories. Treat workers as trusted extensions of the operator machine: use dedicated, disposable worker boxes rather than shared personal or production hosts.

The default worker invocation uses --dangerously-skip-permissions, which removes interactive approval boundaries for autonomous work. Run it only on those dedicated worker boxes, restrict their repository access, and never place credentials in this repository.

Quick start

bin/cnc teams                 # inventory
bin/cnc status                # fleet health: ssh, claude auth, disk, active goals
bin/cnc accounts              # which subs are logged in (human does: cnc ssh <t> --login)
bin/cnc usage --days 7        # how hard each subscription is being pooled

# stand up a box
bin/cnc provision worker-1    # gh-setup + bootstrap (tools + Chromium) + optimize
bin/cnc sync-repos --all      # put every repo + its .env on every box (mirror the workspace)

# secrets live on the MAIN box only (workers just code):
bin/cnc secrets acme --set-creds acme --sync   # render ~/workspace/acme/.env from its vault
bin/cnc secrets acme --set-password acme       # ← --set-password is an alias of --set-creds
bin/cnc bw acme -- get notes ".env webapp"     # ad-hoc secret pull for a Claude session

# run missions
bin/cnc goal "finish the dashboards — 100% complete" --project acme/webapp
bin/cnc goals                          # flight log
bin/cnc deploy-check acme/webapp       # did it actually ship?

Commands

group command
fleet teams · status · accounts · usage · ssh · exec
provision bootstrap (+ Chromium) · gh-setup · optimize [--persist] · provision (all-in-one) · sync-repos
secrets (main box only) secrets <org> [--set-creds <org> | --set-password <org>] [--sync] · bw <org> [-- <bw args>]
scaffold new-team · new-project
missions goal · goals · projects · deploy-check

Selectors: --all · --pool <p> · --tag <t> · <team-id>…

Releases

Releases are cut by pushing a vX.Y.Z tag; .github/workflows/release.yml does the rest.

  1. Set version in package.json to X.Y.Z.

  2. In CHANGELOG.md, move the Unreleased entries under a new ## [X.Y.Z] - YYYY-MM-DD heading and add its link reference at the bottom.

  3. Check the notes render: bash .github/scripts/changelog-notes.sh vX.Y.Z. To preview from GitHub without publishing, run the release workflow manually (Actions, workflow_dispatch) with the tag.

  4. Merge that change to main, then tag the merge commit and push the tag:

    git tag vX.Y.Z && git push origin vX.Y.Z

The workflow runs the CI gate, fails if the tag does not match package.json or CHANGELOG.md has no non-empty section for it, then creates the GitHub release with that section as its notes and the source tarball plus SHA-256 checksum attached. A tag with a suffix (v1.0.0-rc.1) is published as a pre-release.

Map

where what
fleet/teams/*.yml team inventory — schema in fleet/CLAUDE.md; copy _example.yml
fleet/box/ agents.sh (coding-agent aliases) — pushed to boxes · bw-sync.ts (secrets) — runs on the main controller only
projects/<org>/<repo>/ registry: concise-English README + project.yml + goals/ — schema in projects/CLAUDE.md; see projects/_example/
orgs/<org>/bw.yml per-org Bitwarden connection + secret manifest (non-secret) — schema in orgs/CLAUDE.md; see orgs/_example/
bin/cnc the CLI (Bun, zero deps)
scripts/ fleet/ goal/ projects/ deploy/ + lib/
.claude/ /goal /setup commands + fleet/missions skills
docs/ flows (how /goal runs) · architecture · operations · economics · onboarding · secrets · agents

Principles

  • 1 VPS = 1 team = 1 org. Same layout everywhere: ~/workspace/<org>/<repo>.
  • Mirror everything. Every box has every repo + its .env; the commander round-robins goals by load.
  • Humans do OAuth. Subscription logins can't be automated — cnc ssh <team> --login.
  • No secrets in this repo. Account emails and Bitwarden server URLs are identity, not credentials. Master passwords, API keys, and tokens live only on the main box — see docs/secrets.md.
  • Merged ≠ done. A mission ends when deploy-check is green.

About

⚔️ Command & Control center for a pool of VPSes + Claude subscriptions: coding at super scale. 1 VPS = 1 team; dispatch /goal missions, merge PRs sequentially with claudetm, verify deploys.

Topics

Resources

Stars

3 stars

Watchers

0 watching

Forks

Contributors

Languages