lmctl
A provider-agnostic control plane for teams of AI coding agents.
lmctl.com · Documentation · npm · lmprobe
AI agents shouldn't be locked to one provider, one workflow, or one context window. lmctl is a local-first control plane for running teams of AI coding agents — across providers, with adversarial cross-provider review and durable memory, composed in plain text.
It's not an IDE and not another chatbot. lmctl coordinates the agent CLIs you already use (Claude Code, Codex, Copilot CLI, Qwen Code, Antigravity, and more), so a Claude lead can hand coding to Qwen or Codex and have another provider review it — the reviewer a different provider and model from the author, not the same model in a different hat.
npm install -g @lmctl-ai/lmctl
Why lmctl — the three lock-ins
Every agentic tool wants to be the whole platform. That leaves you with three kinds of lock-in. lmctl is built to remove each one:
- Provider lock-in. Most "AI code review" is really self-review — one model grading its own work in a different hat, which rubber-stamps its own blind spots. lmctl makes review adversarial: the reviewer is a different provider and model from the author, so the check is genuinely independent of the work — different models have different (uncorrelated) blind spots, and that diversity is the point: a varied team catches what one model, or its clones, can't. And "provider-agnostic" is literal — alongside the major CLIs, the first-class CLIs include Claude Code, Codex, Copilot CLI, Qwen Code, and Antigravity, while the OpenCode provider reaches any model, local (Ollama) or remote (DeepSeek, Qwen, OpenRouter, Copilot's GPT/Claude/Gemini, …); pick any collection and put them in one team, not one model at a time.
- Workflow lock-in. When most tools say "multi-agent," one provider auto-spawns the agents and you just watch. lmctl puts you in charge: you divide the work and build the team in plain text (a lead talks to its members; teams connect to other teams), choose which provider and model plays each role — top-tier for design, leaner models for routine coding — and tune how they interact. Agents from different providers, orchestrated by you, not clones of one.
- Context-window lock-in. A bigger window still eventually loses to a long-running project, and a session is bound to one provider and one folder. lmctl spreads work across specialized agents (planning, coding, review) and keeps a provider- and directory-agnostic durable-memory — the team's shared brain. If a session fills up or breaks, start a fresh one; nothing is lost.
How it works
- Teams in plain text. Define a team in a .lmctl file — members, the provider/model each one uses, and the role each member plays. No DSL to learn; it reads like a list.
- durable-memory. Every member, whatever its provider, reads durable-memory/ as shared context. Knowledge survives fresh sessions and swapped-in agents.
- Patterns you can reuse. When a team handoff pattern works, write it down as Lead instructions and durable memory. The current public model is teamfile + members + lmctl chat.
Get started
- Install: npm install -g @lmctl-ai/lmctl
- Follow Install & first run and the other tutorials to define your first team and run it.
- Browse the CLI reference and glossary for the full command and teamfile reference.
About this repository
This repo is the source of lmctl.com — the homepage and the documentation site (plus the runnable workflows/, skills/, and examples/ catalogs hosted on the site). The lmctl CLI itself is distributed on npm as @lmctl-ai/lmctl.
The companion code-search CLI lmprobe is documented at lmctl.com/lmprobe, distributed as @lmctl-ai/lmprobe, and has an agent-facing skill at skills/lmprobe-skill.md. The hosted lmprobe manual is sourced from the site/ directory of the public lmprobe repo, normally checked out at ../lmprobe, and published to the s3://lmctl-website-prod/lmprobe/ prefix via scripts/deploy-lmprobe.sh.
To work on the docs locally and deploy the site, see DEPLOY.md:
npm ci
npm run start
Tools
Static tool guides are published at lmctl.com/tools; authoring and validation details are in tools/README.md. Repository checks compile lmscript scripts against the host release declared in each script, using lmscript 0.2.369 or newer; downloads are SHA-256 verified and cached.
Quickrun samples
The hosted autorun catalog is generated from quickrun/*.lms. lmctl autorun lists and runs its scripts; lmctl quickrun remains as a deprecated alias. Add a description comment and deploy the catalog with scripts/deploy-quickrun.sh to add a task without releasing lmctl again. Hosted hello onboarding is an autorun script at helloexample/hello.lms.
