TeamPilot

简体中文 · Development guide · Architecture & AI: AGENTS.md

TeamPilot is a desktop client based on terminal AI Agent Team. Its centerpiece is team capabilities: assign a model — and even a different CLI — per member for tiered collaboration (save tokens, implement fast, review accurately), plus roles, prompts, skills, and plugins in the GUI—then launch one embedded terminal per member that talks to agents through Claude Code, Codex, opencode, cursor, or flashskyai, locally or over SSH. Projects and sessions attach that team to a repo and conversation.

Two ways to work

Mode	When	What you get
Simple mode (personal)	You just want one agent in a repo	Skip the team roster—launch a single CLI and start chatting. No member list to build.
Team mode	Multi-agent / tiered or mixed-CLI collaboration	Define a team once, then run one terminal per member in parallel. This is TeamPilot's centerpiece (below).

Simple mode is not stripped-down. A personal project still gives you the full per-project config—you just skip the multi-member roster:

• Multiple CLIs and models, switchable. Each CLI (Claude Code, Codex, opencode, cursor, flashskyai) keeps its own provider + model + reasoning-effort in the project, so you can configure several models per tool and switch the active CLI/model without retuning globally—the same tiering benefit as teams, for a single agent.
• Per-project agent, skills, plugins, MCP, and extensions—the personal project carries its own prompt and capability set.
• A permanent built-in Personal assistant project, plus any personal projects you create, all alongside your team projects—start simple and graduate to a team when a task needs it.

Core feature: team configuration

Team configuration is what sets TeamPilot apart from a single terminal and hand-typed flags: define the team first, then work in parallel from the chat workbench.

Piece	Purpose
Team	A full multi-agent preset: pick a coordination mode (single-CLI native team, or mixed cross-CLI), team-level CLI options, and the skills/plugins bound to this team only.
Member	A role inside the team (e.g. team-lead, developer, reviewer): its own model, provider, system prompt, launch flags—and, in mixed mode, its own CLI. Connecting a session spawns a separate PTY per member—models and context do not mix.
Skills / plugins	Capabilities attached per team; at launch they are written into an isolated CLI config tree that member terminals inherit.

Per-member model tiers: save tokens, split the work

Using one model for everything either burns premium tokens on trivial edits or underpowers planning and cross-checking. TeamPilot lets each member run a different model tier in parallel—mixing capability, speed, and cost in the same team:

Role (example)	Typical tier	Good for
Lead / planning	High (e.g. Opus, flagship)	Requirements, design docs, scope and acceptance criteria
Implementation	Fast / light (e.g. Haiku, small models)	Bulk edits, scaffolding, getting the happy path working
Review	Mid (e.g. Sonnet)	Code review, gap checks against the plan, cross-file consistency

This is not coding-only: docs, research, ops triage, or any multi-step pipeline can use the same pattern—strong models to think and specify, light models to execute quickly, mid-tier models to verify and align cross-cutting constraints—so you spend less on tokens, ship faster, and hit complex cross-module / cross-functional goals without one chat playing architect, worker, and QA at once.

Typical workflows:

• Tiered models: Set different providers/models for team-lead, implementer, and reviewer; switch member tabs to switch terminal and model—no global retuning each time.
• Parallel roles: team-lead coordinates and delegates (Claude Code expects a member named exactly team-lead); other members handle implementation, review, etc.—switch terminals in one window.
• Mixed CLIs: In a mixed team, members can run different CLIs (Claude Code, Codex, opencode, cursor, flashskyai) and still coordinate through an in-process team bus—use each tool where it's strongest.
• Scenario presets: Maintain teams for “daily dev”, “deep refactor”, “docs”—switch teams instead of retyping models and prompts. Browse and import shareable team templates from the built-in Team Hub.
• Session binding: Opening a project session injects the active team into CLI args (e.g. --team-name / per-member session id, per-member CONFIG_DIR) and can resume prior CLI sessions.

Configure under Settings → Team configuration (route /team-config). Team JSON lives under teams/ in app data; per-member runtime CLI dirs under config-profiles/teams/<team>/members/….

Why TeamPilot?

Skills & plugins (team-scoped)

• Skills: Browse, install, and enable in one place; when mounted on a team, every member terminal in that team shares the same capabilities.
• Plugins: Visual install and management; agree per team which extensions a project uses so environments stay aligned.

Chat workbench

• Multi-tab terminal: Several members and/or sessions in one window instead of many OS terminal tabs.
• Projects & sessions: Organize by repo and chat, with the selected team bound to that work.
• Auto session titles: See what each chat is about in the sidebar.
• Right tools panel: File tree, member list, and prompts next to the chat with less context switching.

Settings & integrations

• RTK (optional): Can be enabled in settings to reduce session overhead and stretch usable context.

Installation

Open the latest GitHub Release and download the asset for your system (names look like teampilot-<version>-…).

Linux

Debian / Ubuntu (.deb, recommended)

sudo dpkg -i teampilot-*-linux.deb
# If dependencies are missing:
sudo apt install -f

Launch TeamPilot from the app menu. Uninstall: sudo apt remove flashskyai-client (exact package name is in the deb metadata).

AppImage (portable)

chmod +x teampilot-*-linux.AppImage
./teampilot-*-linux.AppImage

Requires libfuse2 on many distros (sudo apt install libfuse2 on Ubuntu 22.04+). For menu/Dock integration, use AppImageLauncher.

On desktop, agents run in a local PTY by default. You can switch to SSH in settings so the CLI runs on a remote host.

macOS

1. Download teampilot-*-macos.dmg.
2. Open the DMG and drag TeamPilot into Applications.
3. If Gatekeeper blocks the first launch: allow it under System Settings → Privacy & Security, or right‑click the app → Open.

Windows

Pick one artifact from the same release:

File	Notes
*-windows-setup.exe	Recommended: Inno Setup installer with shortcuts
*.msix	For sideloading / managed deployment
*.zip	Portable: extract and run TeamPilot.exe

If your CLI lives in WSL, point app data or the CLI path at WSL in settings. SSH to a remote Linux dev box is also supported.

Android

Android does not run a local PTY. You connect over SSH to a machine that already has your agent CLI installed.

1. Download teampilot-*-arm64-v8a.apk (most phones) or teampilot-*-armeabi-v7a.apk.
2. Allow installs from unknown sources, then install the APK.
3. In Settings, configure SSH host, user, and key (or password).
4. On the remote host, ensure the CLI is installed and works in the shell you get after SSH login.

Supported CLIs

CLI	Terminal sessions	Provider config	Notes
Claude Code	Yes	Yes	Default team CLI; onboarding can detect/install.
Codex	Yes	Yes	Launchable; joins mixed teams via the team bus.
opencode	Yes	Yes	Config via OPENCODE_CONFIG_DIR.
cursor	Yes	Yes	cursor-agent; HOME-isolated per member.
flashskyai	Yes	Yes	Path resolved at startup.

Before you start

After installation, on the machine where agents actually run (local desktop, or the SSH host on Android):

Item	Notes
Your agent CLI	Whichever CLI your team uses (Claude Code, Codex, opencode, cursor, flashskyai) on the login shell PATH, or set the CLI path under Settings → Session

First launch can run the built-in CLI detection. Installers are built by CI; building from source: Development guide.

More documentation

Doc	Audience	Topic
Development guide	Contributors / maintainers	Setup, run, test, package, CI
AGENTS.md	Contributors / AI	Repo layout, data dirs, architecture

Terminal

Embedded terminals use flutter_alacritty — a Flutter widget backed by an Alacritty-based Rust engine .

Acknowledgements

• File icons: Material Icon Theme (MIT) by Philipp Kief / material-extensions.

License

This project is licensed under the MIT License.