Agents

Orchestrate Claude, Codex, Antigravity, and Cursor in parallel — from one IDE. Open source. Free.

Technical design: AGI EXT architecture.

Turn your editor into a command center for orchestrating Claude, Codex, Antigravity, and Cursor in parallel. Each agent runs as a full-screen editor tab. Each agent can spawn sub-agents. You orchestrate — approving plans, monitoring execution, shipping faster.

Why an IAE?

Text editors became IDEs when coding got complex. Now that AI agents do the coding, your environment needs to evolve again. When you're running 10+ orchestrators each spinning 10+ agents, terminals and TUIs collapse. You need an environment built for that scale.

This extension turns your IDE into an IAE:

• Full-screen terminals - Agents run as editor tabs, not buried panels. See your code and agent side-by-side.
• Session persistence - Close VS Code, reopen it, your agent tabs come back exactly where you left off.
• Approval gates - You approve plans before agents execute. Control without bottleneck.
• Keyboard-first - 12+ shortcuts for spawning, switching, labeling, and reviewing agents. No mouse needed.

Workflow: Task, Plan, Approve

1. Task - Describe what you need with /swarm inside the IDE.
2. Plan - The orchestrator drafts a distribution: who codes, who debugs, who researches.
3. Approve - You gate execution. Agents run only after your approval.

Quick Start

1. Install the extension from VS Code Marketplace
2. Press Cmd+Shift+A to spawn your first agent
3. Open the Dashboard (Cmd+Shift+F) to configure auto-start

agents-dbg Mac App

Install the private Mac app through the Homebrew tap:

brew install --cask muqsitnawaz/tap/agents-dbg

The public install script uses that tap and launches the installed app:

curl -fsSL https://raw.githubusercontent.com/phnx-labs/agents-cli/main/scripts/install-agents-dbg.sh | sh

Navigation

Features

Agent Terminals

Spawn any agent as a full-screen editor tab. Built-in support for Claude Code, Codex, Antigravity, OpenCode, and Cursor. Add custom agents through settings.

Where a new agent runs is agents.launch.defaultTarget: auto (default — agents-cli picks the device), local (this machine), or ask (prompt every time). Under auto, mark your compute boxes once with agents devices role <name> worker and every Agents: New <Harness> rotates over those workers instead of the machine you are sitting at. The device and the version/account are both chosen for you — the account by balanced rotation, which skips signed-out and rate-limited ones — so the command opens an agent without a prompt. Ordinary New, Auto, Pick Host, and Pick Account launches preserve the exact open workspace as their working directory: opening the monorepo root lands at the root locally or remotely, while opening one nested project folder lands in that same nested folder. Local and version-pinned launches preserve this directory too. Agents: New <Harness> (Pick Host) and Agents: New <Harness> (Pick Account) are the orthogonal explicit routes — each pins one placement axis and lets the CLI balance the other. Pick Host asks for the device, then balances the account on it (routing around signed-out or throttled accounts by usage headroom). Pick Account pins the account and auto-picks the device: it auto-selects a worker that can run the interactive account picker and shows that device's usage-aware picker there, so you choose the account — including a signed-out one, which authenticates in place — while the CLI chooses the machine. Agents: New <Harness> (Auto) chooses both automatically, ignoring the configured target. Agents: New <Harness> (Pick Project) never asks for a host — it picks the device and account automatically, like (Auto) — but first shows a quickpick of your projects and lands the agent in the chosen project's path on the auto-selected worker. It is the only launch variant that replaces the workspace with a managed project's configured path.

Session Persistence

Every open agent terminal is fully restorable. Session ID, icon, and custom labels are saved to disk in real-time. VS Code crashes? Restart? All your agent tabs come back exactly as they were.

Automatic tab titles follow the canonical session name from the fleet stream. A remote tab may briefly show its first meaningful prompt, then replaces it when the harness generates a name or the session is renamed; a label set with Cmd+Shift+L remains authoritative. A session dispatched to another box keeps a clean title (CC, or its real name) — the executing device is shown in the status bar instead, shortened to its unique suffix (m3) when fleet peers share a common prefix (yosemite-) — and left whole whenever shortening would stop naming exactly one box.

An Agents terminal can also be reused: exit one agent and launch another in the same tab. The pushed session stream rebinds that tab once the new session is its sole live claimant, so the status bar and Active Session sidebar follow it without a tab switch. Overlapping live claims remain unresolved until the stream removes the old one; start time is not treated as terminal ownership. The join remains scoped to the terminal's observer device and execution host, and the extension does not poll or match raw terminal IDs across machines. This rebind path reads no local session-state file.

Task Management

• Labels - Tag agents by task (Cmd+Shift+L). Status bar shows the active agent, its label, and — for a session running on another device — that device.
• Tasks - Renders tasks returned by agi-cli; the extension does not parse task files or query trackers itself.
• Session history - Browse CLI-provided sessions from the dashboard and resume through agents sessions resume.
• Fork a session - Agents: Fork, Agents: Fork (Recap), and Agents: Fork (Pick Host) delegate the copy and identity semantics to agents sessions fork <id>, which resolves the source across the fleet and launches a recap-seeded same-harness sibling. The original is left untouched; Fork and Fork (Recap) behave identically since the verb is always recap-seeded.
• Fork pairs in the Recap ledger - a fork and the session it came from finish as two rows that share no id. AGI EXT remembers the edge and reunites them: one side-by-side row in Recap, parent on the left, fork on the right, each stamped with the machine it ran on and its own duration/cost/PR.
• Agents: Resume - The picker performs one bounded CLI listing when opened, then sends only the chosen canonical IDs to agents sessions resume <id> --vscodium.
• Resume variants - Agents: Resume (Pick Session) lists only abandoned sessions — detached, backgrounded, parked, or idle, nothing currently open anywhere — and resumes each on the device it was created on. Agents: Resume (Pick Host) reopens the active tab's session on a device you pick (same harness, same version). Agents: Resume (Pick Harness) continues the active tab's session in a different harness on the same device, replaying the transcript through the universal /continue flow. Agents: Resume (Best Profile) rotates the active tab to the signed-in account with the most usage headroom (Cmd+Shift+Alt+J).

Active Session sidebar

The activity-bar sidebar is the session preview: what you asked, and what the agent has actually done about it. It follows the last of two things — the agent terminal you focus, or the row you click in Fleet — so it works for a session with no tab in this window at all.

Its header carries Focus and Resume. Below that: ◈ Goal when the summarizer has one; Request, the user's own sentence in bold with images, files, @dirs, pasted terminal and slash commands lifted out into chips and the whole turn behind a "Full request" fold; Timeline, a step per thing the agent narrated with its tool counts folded in — newest five, three more on demand, one counter for everything older, an amber now-line on the step running right now, and failures, blocks and milestones called out; the Checklist; created plans/artifacts, ticket and PR links, spawned-team summary and current activity; Files the session created, modified or deleted; and a reply field.

Nothing is invented. On an older CLI the request falls back to the cleaned prompt and then the topic, the timeline falls back to the summarizer's checkpoints, and the Files card simply does not appear. When the session stream is not running the panel says so and names the command it expects, and a harness whose transcript cannot be folded says that too — neither renders as a confident empty card.

Generic slash-command buttons, saved prompts, duplicate worktree/CWD navigation, and the separate Issues view are deliberately absent; those belong to their owning command, dashboard, Source Control, or tracker surfaces. All live facts ride the CLI session stream, so switching tabs does not start a transcript scan or poll.

Fleet

The dashboard's mission control. Fleet opens on the session list with All · Active · Idle · Blocked · Favorites, search, sort, and a header-level Dispatch action. The status chips and Group by Status share the same CLI-projected phase/attention taxonomy; grouping by Project, Device, Status, or None re-buckets the entire list, including Previous sessions. Previous history defaults to at most 50 sessions active in the last seven days and pages older history ten rows at a time; bulk Resume applies only to shown resumable rows. Returning after a longer gap — nothing resumable in the last seven days — shows up to ten most-recently-active Previous sessions instead of an empty band, so the list never looks broken; "Show more" still pages further back from there.

A Project control sits beside Group and Sort: a searchable list of every project in view with its session count, plus All projects and Unassigned as first-class picks. Choosing one scopes the rows and the status-chip counts to it, and the pick is remembered. A session's project comes from the hints its stream row already carries — working directory or worktree path, the CLI's repo word, the PR's owner/repo slug, then branch, worktree slug, or team — so cloud and remote runs are attributed without a local checkout. Runs that genuinely have no repo land in a labeled Unassigned group that states why (“cloud & no-repo runs — no cwd to match”) instead of an empty header, with + assign project on each row and assign all on the group header to file them by hand.

Each row uses the harness logo without repeating its name. Its chevron expands the row in place, showing the CLI-indexed full first user turn as Your request (with See more for long requests), or labels an older CLI's topic fallback honestly as Request summary — independent of clicking the row itself, which points the Active Session sidebar at that session. Ended-session timing reads “Ran … · ended … ago” with the absolute range on hover, and names its device once as “Runs on …”. Ticket, PR, and repository pills navigate; branch pills copy. The single compact Resume targets the owning device, while right-clicking it offers Resume on a different device… from the already-loaded fleet roster; that path delegates transcript transport and liveness to agents sessions migrate, then focuses the target in the same editor terminal. Selecting a session sends it to the Active Session sidebar — Fleet has one job, listing sessions, and the sidebar has the other, describing one. Background/headless runs remain hidden by default in the agent feed; use the Background toggle when you need them.

Needs you projects the CLI feed's canonical attention ledger. Cards retain the exact choices and source; answered items leave the count immediately while the activity lane keeps a compact resolution receipt. The extension does not inspect transcripts or independently poll GitHub to decide what needs a person.

Agent cards surface outputs such as PRs, spawned teams, created tickets, and plan artifacts. .html and ref-*.md plans detected in session output, worktree files, or attachments appear as one-click preview chips. Screenshots and other session attachments appear as thumbnail artifacts that click through to the host preview. The expanded project and host sidebar can be resized by dragging its right edge.

Selecting a session costs nothing: the sidebar renders it from the session stream the extension already holds, so a click starts no CLI process at all.

Foreman Voice Orb

Talk to your fleet. Tap the orb in the dashboard and ask "what's running?", "what's left this cycle?", or "dispatch RUSH-557 to the cloud" — a realtime voice model answers out loud with live floor state, and it can dispatch tickets, spawn agents, and file Linear issues for you. Tap to talk, tap again to stop, or press-and-hold to talk only while pressed. A silent-mode toggle under the orb switches replies to text-only transcript.

AI Git Commits

Generate commit messages from staged changes with Cmd+Shift+G. Learns from your commit style, then stages, commits, and pushes in one action.

Additional Features

• Auto-start - Configure which agents launch when VS Code opens
• Default models - Set preferred model per agent type
• Shell terminals - Spawn plain shells alongside agents (Cmd+Shift+S)
• Reader - Notion-style .md editor (TipTap) plus sandboxed HTML preview for artifacts-cli pages; Command Palette: Agents: Reader (Enable) / Agents: Reader (Disable)
• Notifications - Native macOS notifications when agents need attention

For Teams

• Shared dashboard keeps approvals and assignments visible.
• Consistent agent distribution prevents over- or under-staffing tasks.
• Approval gates make it easy for leads to review before code runs.

Requirements

• VS Code or Cursor
• Agent CLIs installed (claude, codex, antigravity, cursor-agent, opencode)
• OpenAI API key (optional, for commit generation and the Foreman voice orb)
• ffmpeg (optional, for the Foreman voice orb: brew install ffmpeg)

Related Packages

• @swarmify/agents-cli - Moved to its own repository

License

MIT