| name | agents-cli |
| description | Manage AI coding agent CLIs with agents-cli. Triggers on: 'agents add', 'agents use', 'agents pull', 'agents push', installing agent versions, syncing agent config, switching between agent versions, managing MCP servers, or when user mentions the agents-cli tool. |
| author | phnx-labs |
| version | 2.0.0 |
agents-cli
agents is the control plane for AI coding agents on this machine. You have it available right now. It can do far more than version management — use it to recall prior work, coordinate parallel agents, dispatch to cloud, and drive interactive terminals.
When you need exact flags or syntax for anything below, run agents <command> --help directly.
Session Recall
The most underused capability. Every conversation you (and other agents) have had is stored and searchable.
You can search across sessions by topic, file path, or free text. If you're about to implement something and wonder whether another agent already worked on it, search first.
You can filter sessions by agent, project, or time window. Useful when you want to see only Claude sessions from the last week, or all sessions touching a specific repo.
You can extract specific roles from a session. Pull only the user's messages to understand what was asked. Pull only thinking blocks to see the reasoning chain. Pull only tool calls to audit what was done. The --include and --exclude flags control this, and they compose with --first/--last to get a slice of the conversation.
You can pull artifacts. Every file written or edited during a session is tracked. You can list them or read a specific one by name — useful to recover work without re-reading the whole session.
You can check what's running right now. --active shows live sessions across terminals, teams, cloud, and headless agents. Use this before spawning duplicate work.
You can read cloud-run sessions. Sessions from cloud providers are also queryable via --cloud.
The key insight: sessions are your memory. Before starting a task, search for prior sessions on the same topic or repo. Before spawning a subagent, check if one is already active.
Running Agents Programmatically
agents run <agent> "prompt" executes an agent headlessly and returns when done. You can set the reasoning effort, working directory, mode (plan/edit/full), and inject secrets or env vars. You can also resume a previous Claude session by session ID.
Use this when you want to delegate a bounded task to another agent and capture its output, rather than spawning a full team.
Teams: Parallel Multi-Agent Coordination
agents teams lets you create named groups of agents working in parallel on a shared task, with optional DAG-style dependencies (--after). Each teammate runs in the background; you use status to check in.
Team sessions appear in agents sessions --teams so you can read what each teammate did.
Use teams when a task has independent parallel workstreams (backend + frontend) or sequential dependencies (QA waits for both to finish). Run agents teams --help for the full interface.
Cloud Dispatch
agents cloud run dispatches a task to a remote cloud agent. The task runs against a GitHub repo and branch. You can follow logs live, send follow-up messages, or cancel.
Use cloud dispatch when the task is too long-running for a local session, needs a clean environment, or should produce a PR.
PTY: Interactive Terminal Sessions
agents pty gives you a real terminal session — essential for REPLs, TUIs, interactive CLIs, or anything that won't work in a plain shell. Start a session, send commands non-blocking, read the screen as clean text, send keystrokes.
You can also use agents pty to drive the agents CLI itself from within another agent — useful when you need to launch an interactive picker.
Routines: Scheduled Recurring Agents
agents routines schedules agents to run on a cron schedule. Check existing routines before setting up new ones. You can run a routine immediately in the foreground to test it before enabling the schedule.
Version and Config Management
Install, switch, and inspect agent CLI versions. agents view shows installed versions and resources. agents view <agent>@<version> shows exactly what commands, skills, MCP servers, and memory files are active for a version.
Config (commands, skills, hooks, memory, MCP) is git-tracked in ~/.agents/ and syncs across machines via agents pull / agents push.
