rocket

Quick start · Website · How it works · AGENTS.md · llms.txt · Skills · Discussions · All projects

Release engineering you can put in CI.

Rocket is an agent employee, a packaged specialist agent with its own skills, memory and tests. It owns one job: deploy and CI. Coding agents use it through its skills, its MCP server and its command line. A scheduled job runs its evals and proposes fixes. A person makes every change.

Rocket is part of an agent-native play for agent-native systems that can eventually be recursively self-improving and evolving, operated by agent employees.

rocket is the deploy and CI/CD employee for coding agents: a senior release engineer that gets your code to production, proves it got there, and keeps rollback possible. It treats a green checkmark as the most dangerous lie in its field, because a green check binds to a commit, an artifact and a config, and rocket verifies all three.

Tagline: "Ships it. Proves it shipped. Rolls it back in seconds." Powered by Energy.

Use it from any coding agent

Claude Code: the plugin brings the skills, sub-agents, hooks and MCP server.

/plugin marketplace add naman10parikh/rocket
/plugin install rocket@rocket

Its MCP server, Codex, Cursor, Gemini CLI, one skill and the command line

Its MCP server, in any MCP client:

claude mcp add --transport stdio rocket -- npx -y -p github:naman10parikh/rocket rocket-mcp
codex mcp add rocket -- npx -y -p github:naman10parikh/rocket rocket-mcp

Cursor (.cursor/mcp.json), Gemini CLI (.gemini/settings.json) and most other clients take the same command as JSON:

{
  "mcpServers": {
    "rocket": {
      "command": "npx",
      "args": [
        "-y",
        "-p",
        "github:naman10parikh/rocket",
        "rocket-mcp"
      ]
    }
  }
}

Gemini CLI: the extension brings the context and the MCP server.

gemini extensions install https://github.com/naman10parikh/rocket

One skill (5 in all, in skills/), in any agent that reads Agent Skills:

gemini skills install https://github.com/naman10parikh/rocket --path skills/deploy-verify
git clone --depth 1 https://github.com/naman10parikh/rocket /tmp/rocket && cp -r /tmp/rocket/skills/deploy-verify ~/.claude/skills/

The command line:

npx -y -p github:naman10parikh/rocket rocket --help

Codex, Cursor, OpenCode, GitHub Copilot and other agents read AGENTS.md; Claude Code reads CLAUDE.md and Gemini CLI reads GEMINI.md, which import it. llms.txt maps every surface for agents that read first.

Quick start

Run it straight from GitHub, with no clone:

npx -y -p github:naman10parikh/rocket rocket --help
npx -y -p github:naman10parikh/rocket rocket --version

From a clone (Node 18 or later; the CLI's build is committed in tools/rocket/dist/):

cd tools/rocket && pnpm install --frozen-lockfile && pnpm build && cd ../..
node tools/rocket/dist/cli.js --version
node tools/rocket/dist/cli.js wire --target ./fixtures/sample-next-app --dry-run
node tools/rocket/dist/cli.js audit --json ./fixtures/bad-workflow.yml
(cd tools/rocket && pnpm test)   # 55 tests
npm run eval                     # golden tasks: 4 at L1 and 6 at L2

wire --dry-run prints the plan (detect, generate, validate, commit, verify). audit reports four findings in the sample workflow, one of them P0, and says to fix the P0 before any deploy. A local cockpit shows the brain, the run log, the eval scores and an ask-the-employee box: bash scripts/serve.sh, then open http://localhost:7799.

How it works

Point rocket at a repository and it wires a lint-clean GitHub Actions pipeline, a reachable deploy and a preview on the second commit (rocket wire). Ask it to prove a deploy is real and it checks both axes: the right artifact is live, and the live artifact behaves. It answers GO or NO-GO with the gauges that proved it. Its moat is that verify-don't-trust discipline and eval-gated CI (block on regression, not an absolute floor), more than the workflow generator.

Command	MCP tool	What it does
rocket wire --target <repo> [--lane vercel] [--dry-run]	rocket_wire	Generates an actionlint-clean workflow and Vercel deploy config, and names the two files for you to commit. The Fly, Cloudflare and Railway lanes are planned. This is the acceptance test.
rocket audit <workflows dir or repo>	rocket_audit	Sweeps the supply-chain attack classes with actionlint and zizmor; P0 and P1 findings with SHA-pin fixes.
rocket verify --target <repo|url> [--sha <expected>]	rocket_verify	The both-axes check: deployed SHA equals HEAD, and a behaviour smoke test passes. Answers GO or NO-GO.
rocket rollback --target <repo> [--to <deployment|image>]	rocket_rollback	The rollback hierarchy, fastest first: flag, alias repoint, image pin, git revert.
rocket eval-gate --baseline <ref> [--floor <n>]	rocket_eval_gate	Gates a release on regression against a rolling baseline plus an absolute floor.
rocket migrate-plan --diff <schema-diff>	rocket_migrate_plan	The expand-contract three-phase plan; refuses a destructive change coupled with dependent code. Plan only.
rocket preview reap --repo <repo> [--older-than <d>]	rocket_preview_reap	Lists, then tears down, leaked preview environments.
rocket doctor		Checks the toolbelt: act, actionlint, zizmor, gh, secrets, a searchable brain, a green eval.

Every read path takes --json. Exit 0 means clean; non-zero means a finding, a NO-GO or a refusal. Each MCP tool is a thin wrapper over the same core the CLI calls, so the two never drift. The plugin brings the skills, the two sub-agents (a pipeline auditor and a deploy verifier), the house-law hooks and the MCP server.

The three house laws (from brain/OPINIONS.md):

1. Deploy strategy: blue-green by default; canary only when telemetry exists. For AI agents the canary metric is the eval score.
2. Eval-gated CI: block on regression, not an absolute floor (a dual threshold: a floor plus a Welch's t delta against a rolling baseline). The gate is the evolving golden dataset, with no self-grading.
3. Vercel: one push at a time, through Git integration only, with "Require Verified Commits" off, no deploy-hook workflows and no manual vercel --prod. rocket's PreToolUse hook enforces this.

The boundary. rocket owns the pipeline and the deploy, end to end, and only that. Product features go to pixel (frontend), turbo (backend) and pebble (mobile); secrets at rest to hush; error monitoring to canary; the schema to acorn; auth and payments to badge and penny; the earning loop to helios. rocket plans the migration ordering (expand-contract) and wires the siblings' deploys; it does not write their code.

Where to read next: AGENTS.md is the agent's entry point (every tool, the boundary, the NOT-list), llms.txt maps every surface, and CLAUDE.md is the Claude Code operating doc. The six expertise files in brain/ (jobs, scenarios, pitfalls, opinions, gaps, tool matrix) answer to bash scripts/memory-search.sh "<question>". npm publishing is parked; the CLI installs from GitHub.

What it is made of

Every project in this fleet is graded on the same 11 harness components and on self-improvement. Each mark is backed by a file you can open.

Component	Grade	Where
Identity	✅	identity/SOUL.md:5
Memory	✅	memory/MEMORY.md:1 + scripts/memory-search.sh:1
Skills	✅	skills/deploy-verify/SKILL.md:1
Tools/CLIs	✅	package.json:7 + tools/rocket/src/cli.ts:1
Hooks	✅	.claude/settings.json:4 + hooks/hooks.json:4
Sub-agents	✅	.claude/agents/pipeline-auditor.md:1
MCPs	✅	tools/rocket/src/mcp-server.ts:113
Dispatch	✅	tools/rocket/src/routing.ts:55 + .github/workflows/rsi.yml:12
Evals	✅	eval/run.mjs:1 + .github/workflows/ci.yml:60
Observability	✅	tools/rocket/src/log-run.ts:77
Sandbox/Infra	✅	tools/rocket/src/sandbox/e2b-runner.ts:104
Self-improvement	✅	dsip.json:1 + .github/workflows/rsi.yml:46

✅ real and tested · 🟡 partial, with the reason in the row · ❌ absent

How it improves itself

It carries its own evals and a self-improvement contract, dsip.json, which says what it may change and how a change is judged. A scheduled job runs its evals and proposes fixes. A person makes every change. Lessons it learns are kept as rules in its memory files.

Part of a fleet

One of the projects by Naman Parikh. Each owns one part of an agent's harness, or of the product around an agent, and all of them follow one standard.

Related project	What it owns
AgentDial	Identity
WikiMem	Memory
Skillsmith	Skills
Agent Booster	Tools / CLIs
Hooksmith	Hooks
AgentCast	Sub-agents
Plugsmith	Plugins
Armory	MCPs and the component registry
AgentGrid	Dispatch
AgentSwarm	Dispatch
AgentBench	Evals
AgentMoney	Observability
Sandforge	Sandbox / Infra
Claude Harness	The whole harness, as a template

Community

Questions and ideas go to Discussions; bugs go to issues. Follow the work on X and LinkedIn.

Contributing and security

Read CONTRIBUTING.md. AI-assisted pull requests are welcome. Report a vulnerability privately, as SECURITY.md explains.

Star history

If it is useful to you, a star helps other people find it.

Contributors

License

MIT. See LICENSE.