Dispatch Claude Code, Codex, Cursor, and Grok across your own machines and subscriptions. Measure every run, fold the lesson back into the harness, and schedule what should run itself — a fleet you steer, not another chat window.
Claude Code is one agent. agi-cli runs the whole factory.
also available as ag.
Built by engineers from
Real parallelism, not a metaphor: agents work at the same time, each version-pinned per repo, on subscriptions you already pay for. One engineer steers; the factory plans, codes, tests, and ships — then you measure, encode the lesson, and put the loop on a schedule.
agents perf and agents insights show what's slow and what breaks; fold the fix into AGENTS.md or a skill and every agent picks it up next run.Profiles resolve model, provider, version, and credentials before the underlying CLI starts.
One prompt surface for the CLIs you already use. Pin versions per project, swap models through profiles, rotate accounts when quota is tight, and keep the command shape stable. Run docs →
$ agents profiles add kimi $ agents run kimi "refactor the queue worker"
The harness keeps quota, model choice, and project defaults out of ad-hoc shell glue. Runs spread across your accounts with the balanced strategy, one agent pipes into another, and the project decides the installed version. Rotation docs →
$ agents run claude --balanced "run the full test suite" $ agents run claude "Find auth bugs" | agents run codex "Patch them" $ agents usage
One agent is the on-ramp. Parallel work gets named teammates, boundary contracts, DAG dependencies, logs, and cloud dispatch — on one machine or across your fleet.
Teams attempt every ready DAG node independently. Permanent launch failures keep structured evidence and block only descendants, while independent branches continue; worktrees are opt-in with --enable-worktrees. Teams docs →
$ agents teams create checkout $ agents teams add checkout codex "Owns: app/payments/*" --name payments $ agents teams add checkout claude "Run Playwright" --name qa --after payments $ agents teams start checkout --watch
Teammates run on other machines over SSH. Declare a device pool and the scheduler places them, or pin a teammate to a specific box; agents doctor --devices shows agent readiness across the fleet. Teams docs →
$ agents teams create feat --devices yosemite-s0,yosemite-s1 $ agents teams add feat claude "Owns: api/*" --name api $ agents teams start feat --watch $ agents doctor --devices
One config source syncs skills, MCP servers, commands, hooks, and rules into every harness's native format — edit once, every agent sees it.
Skills, MCP servers, slash commands, hooks, and permissions install into ~/.agents/ and sync to every active agent version — no more claude mcp add, then codex mcp add, then editing Gemini's config by hand. Sync docs →
$ agents skills add gh:yourname/python-expert $ agents install mcp:com.notion/mcp $ agents commands add gh:yourname/commands
Write one AGENTS.md and it becomes CLAUDE.md, GEMINI.md, and .cursorrules. Resources stack in four layers — system, you, team, project — and an agent sees the merged surface at spawn, highest layer winning. Config layers →
# ~/.agents/ is the canonical source ▒ system built-in skills + commands ▒ you ~/.agents/ ▒ team .agents/ (shared repo) ▒ project ./.agents/ $ agents doctor
Running agents aren't fire-and-forget. Steer them mid-flight, unstick what stalls, and make every machine match — a 1,100-agent fleet reads as dozens of deliverables.
agents feed collapses the fleet under the outcome each agent serves — a Linear ticket, a PR, a worktree. agents message <target> reaches any live run, teammate, or cloud task by name, and the text lands at its next tool call. Fleet docs →
$ agents feed $ agents message payments "switch to the v2 schema"
agents watchdog detects a stalled session, resolves the exact terminal split it lives in — tmux, iTerm, VSCodium, or a raw pty — and injects a nudge. The fleet keeps moving without you babysitting it. Watchdog docs →
Claude Code
Codex
Cursor
OpenCode
Copilot$ agents watchdogOne machine is set up the way you like it; agents apply installs missing agents, upgrades the CLI, syncs config, and propagates logins — turning six hosts × eight harnesses = 48 OAuth flows into one. Fleet sync docs →
$ agents devices list $ agents apply
Some work shouldn't tie up your laptop. Hand it to a managed runner that clones the repo, plans, implements, tests, and opens a PR.
agents cloud run sends a task to Rush Cloud, Codex Cloud, a Factory droid, or Antigravity. It works the repo end-to-end and opens a pull request while your terminal stays free. Cloud docs →
$ agents cloud run "add rate limiting to the API" --provider rush $ agents cloud list
Register a run as an event trigger with --on pull_request — also push, issue_comment, and workflow_run. Your repo gets an agent that reacts, with no CI glue to maintain. Event triggers →
$ agents cloud run "review this PR" --provider rush --on pull_request
Every transcript lands in one searchable store, so prior attempts — and prior spend — are a command away.
Find the debugging trail, the abandoned attempt, or the one command that worked last week. Sessions can be searched, replayed as markdown, filtered by project, and tailed while agents are still running — and sessions focus drops you back into any of them, wherever it lives. Sessions docs →
$ agents sessions "stripe webhook signature" $ agents sessions a7f3e2c1 --markdown $ agents sessions --active $ agents sessions focus a7f3e2c1
Every session is priced at scan time from an offline model-price table. agents cost rolls up $ and duration across every account and harness; agents budget caps a run before it starts and kills a headless run the moment it crosses the line. Costs docs →
$ agents cost --since 30d --by project $ agents budget set per_run 5 $ agents insights --by account
Package repeatable orchestration and executable surfaces into named, portable bundles that resolve project → user → system.
Workflows bundle an orchestrator prompt with subagents, skills, and plugins into one named invocation. WORKFLOW.md is the orchestrator's system prompt; files under subagents/ dispatch by name, in parallel. Workflow docs →
$ agents run deep-research "compare vector databases"
Plugins carry skills, commands, hooks, MCP servers, and permissions under a single manifest, mirrored into every installed version automatically. Switch versions with agents use and the plugin re-syncs — no reinstall. Plugin docs →
$ agents plugins install gh:yourname/security-suite $ agents plugins list
Browser, secrets, routines, PTY, computer, and devices are command surfaces under the same harness.
The harness reaches the parts agents usually lose: logged-in browser profiles, Keychain bundles, scheduled runs, remote devices, native apps, and interactive terminals. agents run --host puts a whole run on another machine over SSH. Devices docs →
$ agents browser start --profile work --url https://linear.app $ agents secrets exec prod -- npm run deploy $ agents pty start -- npm run dev $ agents run claude --host mac-mini
Routines turn repeatable agent work into a schedule. Standups, weekly digests, nightly audits, and one-shot follow-ups stay tied to the same agent profiles and logs. Schedule docs →
$ agents routines add standup \ --schedule "0 9 * * 1-5" \ --agent claude \ --prompt "Draft a standup from yesterday's git log"
On macOS, press Command-Shift-O (⌘⇧O), describe the problem, add screenshots, and choose File Ticket or Fix. A ticket agent carries every selected file into the issue wherever it fits best. Fix sends the same context to every selected agent. Menu bar docs →
⌘⇧O describe checkout email is blank mode File Ticket agent Claude files 3 screenshots attached
The tool is entirely free — the best tools should be in every developer's hands.
The CLI binary has no built-in telemetry or phone-home path.
Run on your own subscription or API key. No account required — an optional sign-in adds team spaces. Still no upsell.
Most tools give you one piece. agents gives you all of them.
| Capability | agi-cli | Claude Code alone | Cursor | Run CLIs by hand | OpenCode |
|---|---|---|---|---|---|
| Pin versions per project (.nvmrc-style) | ✓ | — | — | — | — |
| Run multiple agents (Claude + Codex + Grok) from one CLI | ✓ | — | — | ~ | — |
| Swap underlying model (Kimi, GLM, DeepSeek via OpenRouter) | ✓ | — | ~ | — | ✓ |
| Rotate across multiple accounts to dodge rate limits | ✓ | — | — | — | — |
| Parallel teams with DAG dependencies | ✓ | — | — | — | — |
| Distribute teammates across machines over SSH | ✓ | — | — | ~ | — |
| Steer & auto-nudge running agents mid-flight | ✓ | — | — | — | — |
| Managed cloud run that opens a PR | ✓ | — | ~ | — | — |
| Reusable multi-agent workflow bundles | ✓ | — | — | — | — |
| Local browser via CDP (drive any site) | ✓ | — | — | — | — |
| Cross-agent session search & replay | ✓ | ~ | — | — | ~ |
| Cron / scheduled routines | ✓ | — | — | ~ | — |
| Keychain-backed secrets (no .env files) | ✓ | — | — | — | — |
| Sync skills/MCP/commands across all installed agents | ✓ | ~ | — | — | ~ |
| Source-available and self-hosted — cloud dispatch optional | ✓ | — | — | ✓ | ✓ |
vs: factory-ai · opencode · manual clis
OpenClaw
Kiro
Goose
Gemini$ curl -fsSL agi-cli.sh | shagi-cli (AGI CLI) is an open-source meta-harness that runs every CLI coding agent — Claude Code, Codex, Cursor, Grok, and more — from one interface. Install it with curl -fsSL agi-cli.sh | sh; it lands on your PATH as agents (alias ag). One install, one config, every agent.
Yes — every feature, for every user, forever. It runs agents on your existing subscriptions, so there are no per-token API costs and no account required. An optional sign-in adds team spaces; nothing is gated behind it.
FSL-1.1-Apache-2.0. Use it, modify it, redistribute it — for any user or company, free of charge. The only thing barred is offering agi-cli — or a substitute for it or for other Phoenix Labs products built with it — as a competing commercial product or service. Each version converts to plain Apache-2.0 two years after it ships.
Claude Code, Codex, Cursor, Antigravity, OpenCode, Grok, Kimi, Droid and more — plus any model (Kimi, MiniMax, GLM, Qwen, DeepSeek) through OpenRouter profiles.
Claude Code is one agent. agi-cli runs and orchestrates all of them — version-pinned per project like .nvmrc, in parallel teams, with skills, secrets, and searchable sessions shared across every agent.
Run curl -fsSL agi-cli.sh | sh. Works on macOS, Linux, and Windows.