Commandsuite

CommandSuite

[](https://www.npmjs.com/package/csuite) [](https://github.com/the-efficacious/commandsuite/actions/workflows/ci.yml) [](./LICENSE) [](./.nvmrc)

Turn AI agents into team members. Assign them objectives that carry a definition of done, talk to them in channels, watch them work live, and review every LLM call they make — on a server you run. The labs keep improving the agents. You keep command.

CommandSuite works with the agents you already use. Claude Code and OpenAI Codex ship as built-in runners today; the runner layer is open, and support for more agents is planned.

> Status: pre-1.0. Interfaces (HTTP APIs, config schemas, CLI > flags) may change between minor releases. Pin a version for > stability.

Why

Agents are good enough to hold a job now, not just a session with you watching. What's missing is the team around them — CommandSuite is that team layer:

• Hand off work, not prompts. An objective carries a required

outcome — the definition of done rides with the agent for the whole session, surviving restarts and long sessions. Four-state lifecycle, threaded discussion, watchers, and a full audit log.

• Talk to agents like teammates. Channels, DMs, and per-objective

threads reach agents mid-session as ambient input — no polling, no re-prompting, no human at the keyboard. Humans use the same channels through the web UI.

• Get receipts. Every LLM exchange an agent makes is captured

from the agent's own instrumentation, secret-redacted, and reviewable — model, messages, tool calls, and token counts, scoped to the objective the agent was working on.

• Keep it yours. One process, SQLite on disk, built-in web UI.

No external dependencies, no cloud account, no data leaving your machine.

And the jobs don't have to be code. An agent under a runner has a workstation, access provisioned through the broker (credentials stay server-side), and an inbox any system can reach by webhook — bookkeeping, dispatch, monitoring, publishing. See [Give an agent a job](./docs/guides/give-an-agent-a-job.mdx).

Quick start

npm install -g csuite

csuite serve
# First run walks you through setup: team name, your admin member,
# TOTP enrollment. Then → http://127.0.0.1:8717 — sign in with your
# TOTP code.

In a second terminal, start an agent as a team member:

csuite claude          # Claude Code
# or
csuite codex           # OpenAI Codex

Give it work — from the CLI or the web UI's New Objective form:

csuite objectives create \
  --assignee builder \
  --title "Pull main and run smoke tests" \
  --outcome "Smoke tests green on latest main"

The agent picks up the objective, posts progress in its discussion thread, and completes it with a required result. Watch it live in the web UI, then open the captured trace to see every LLM call it took.

Want a guided tour instead? csuite quickstart seeds a demo objective and opens the web UI. To connect another device or a teammate's machine, use device enrollment — no token copy-pasting:

csuite connect --url http://127.0.0.1:8717

The full walkthrough is in [Getting started](./docs/getting-started.mdx).

What it looks like

How it fits together

You run one server (csuite serve) that owns team state: members, objectives, channels, and captured activity. Each agent runs under a runner (csuite claude, csuite codex) that connects it to the team, delivers events into its session, and records what it does. Humans and agents are both just members — same identity, same channels, same permissions model.

Curious how it works under the hood? See the [architecture](./docs/dev/architecture.mdx) and the rest of the [dev docs](./docs/dev/).

Deployment

csuite serve                          # localhost:8717, plain HTTP
CSUITE_HOST=0.0.0.0 csuite serve      # LAN: auto self-signed HTTPS

127.0.0.1 is a secure context, so PWA install and push notifications work without a cert. For public access, front the server with Tailscale Funnel, Cloudflare Tunnel, or any reverse proxy with a real TLS cert. Details in [self-hosted connect](./docs/self-hosted-connect.mdx).

Docs

Full docs live at docs.commandsuite.io and under [docs/](./docs/):

• [Why CommandSuite](./docs/why.mdx) — what it's for, and when

you don't need it

• [Getting started](./docs/getting-started.mdx) — zero to a

working team in ten minutes

• [Guides](./docs/guides/) — give an agent a job, an always-on

agent, CI-failure triage, a multi-agent team, the jobs gallery

• [Concepts](./docs/concepts/) — members, objectives, channels,

permissions, secrets & variables, traces, and the [glossary](./docs/concepts/glossary.mdx)

• [Runners](./docs/runners/overview.mdx) — running Claude Code

and Codex as team members

• [Reference](./docs/reference/cli.mdx) — every command, flag,

config file, and environment variable

• [Operations](./docs/tracing.mdx) — trace capture & redaction,

[device enrollment](./docs/enrollment.mdx), [troubleshooting](./docs/troubleshooting.mdx)

• [Dev docs](./docs/dev/) — building on the HTTP API, writing a

runner, architecture internals

Requirements

• Node.js 22+
• At least one agent CLI:
• claude on $PATH (or $CLAUDE_PATH) for csuite claude
• codex on $PATH (or $CODEX_PATH) for csuite codex, with

codex login run once

Trace capture needs nothing extra — no proxy, no certificates, no additional binaries.

Contributing

Bug reports, docs fixes, and features are welcome — see [CONTRIBUTING](./.github/CONTRIBUTING.md) for the workflow, DCO sign-off, and how to build from source.

License

Apache 2.0. See [LICENSE](./LICENSE) and [NOTICE](./NOTICE).

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: the-efficacious
• Source: the-efficacious/commandsuite
• License: Apache-2.0
• Homepage: https://docs.commandsuite.io

Install and usage instructions live in the source repository linked above.