Agent Deck
A native Mac workspace that wraps your installed Pi CLI so sessions, skills, prompts and parallel work stay visible and controlled.
- Source repo
- a-streetcoder/agent-deck
- Stars
- ★ 124
- Last updated
- 3mo ago
- License
- MIT
- Primary language
- Swift
- FA score
- 45/100 · Major gaps
At a glance
- How it runs
- Works with
- Platform-specific
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- Developers already using the Pi CLI who want a native Mac window to watch several sessions and subagents at once.
- Not a fit if
- Teams on Windows or Linux — macOS 26 and Apple Silicon only
- Users who don't want the Pi CLI installed — Agent Deck is not a runtime
- Anyone on Intel Macs or older macOS versions
- Website
- agentdeck.site ↗ set on GitHub
- Maintenance
- Commits in the last 6 monthslast commit 2026-08-01latest release v5.1.3 (2026-07-24)
- Source review
- 45/100 · Major gaps
What does this agent do, and when should you use it?
Agent Deck is a native macOS app built with Swift 6 and SwiftUI that acts as a workspace around the Pi coding agent rather than a runtime of its own. It launches the Pi CLI you already have in JSONL RPC mode and surfaces sessions, tool calls, plans, file previews and steering messages in one window. It organizes agents, skills, reusable prompts and slash commands per project or across projects, keeping built-in resources read-only and routing customizations through overrides with write targets shown before work starts. Parallel work is kept tidy with isolated git worktrees per session or subagent, which can be reviewed and merged back on your terms. Distribution is a signed and notarized DMG or an install.sh script, with Sparkle updates, and it requires macOS 26 (Tahoe) on Apple Silicon.
The app launches your installed Pi CLI in JSONL RPC mode and renders a live session: message cards, tool calls, plans, file previews, attachments, steering messages and terminal handoff, plus session-scoped composer drafts. It scans and organizes agents, skills, prompts and slash commands for one project or across projects, showing their scope and applying them only when deliberately assigned; built-in resources stay read-only, customizations use overrides, and write targets are displayed before work begins. App-managed subagents can be delegated research, planning, implementation or review, and report-only subagents write to their own artifacts rather than your project. Sessions and subagents can be given isolated git worktrees, start from a GitHub issue, then be reviewed, committed, pushed or merged. It also offers model and provider browsing, per-session or per-agent preferences, masked global environment keys, project-level Markdown memory with secret scanning on writes, and a Doctor plus onboarding flow to find, install and update Pi.
- Developers already using the Pi CLI who want a native Mac window to watch several sessions and subagents at once.
- Teams splitting research, planning, implementation and review across subagents and tracking each one's output.
- Engineers starting from a GitHub issue and wanting changes made in an isolated worktree before review and merge.
- People maintaining agents, skills, prompts and slash commands across multiple projects who need to know exactly what entered a session.
- Small teams keeping project-specific Markdown memory for decisions and runbooks while worried about leaking secrets.
- Users who want to browse models and providers, switch preferences per session and manage global environment keys in one place.
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Install / run
curl -fsSL https://raw.githubusercontent.com/a-streetcoder/agent-deck/main/install.sh | bash
- Install notes cover
- macOS
How do you install or deploy this agent?
Requires macOS 26 (Tahoe) and Apple Silicon. Install the Pi CLI first (onboarding can do it for you), then pick one path:
curl -fsSL https://raw.githubusercontent.com/a-streetcoder/agent-deck/main/install.sh | bashOr download the signed and notarized DMG from Releases and drag Agent Deck to /Applications. The script lives at install.sh; read it and its history before piping it to your shell.
Building from source requires Xcode 26.4+:
git clone https://github.com/a-streetcoder/agent-deck.git
cd agent-deck
open agent-deck.xcodeprojOr build on the command line:
xcodebuild -project agent-deck.xcodeproj -scheme agent-deck \
-configuration Debug -destination 'platform=macOS' \
CODE_SIGNING_ALLOWED=NO buildHow do you use this agent?
On first launch, use Doctor and onboarding to locate or install the Pi CLI; if Pi is already present, the app launches it in JSONL RPC mode.
Create a Pi session in a project context to see message cards, tool calls, plans and file previews. Assign the agents, skills, prompts and slash commands that session should use — files on disk do not become part of a session until deliberately assigned — and confirm scopes before starting.
For parallel work, give a session or subagent an isolated git worktree, start from a GitHub issue on its own branch, do the work in that directory, then review and commit, push or merge.
Model and provider preferences, masked global environment keys and project-level Markdown memory each have their own views; memory writes are secret-scanned. Production builds send one minimized app_opened event per launch to EU-hosted PostHog; the integration is off in Debug, XCTest, previews and benchmarks.
What are this agent's strengths and limitations?
- Native SwiftUI app with no Electron or embedded web UI, using system shortcuts and native decision prompts.
- Explicit scope and write visibility: built-in resources are read-only, customizations use overrides, write targets appear before work starts.
- Report-only subagents write to their own artifacts instead of your project directory.
- Sessions and subagents can work in isolated git worktrees and be merged back on your schedule.
- Signed, notarized releases with standard macOS updates via Sparkle.
- MIT licensed, with source and detailed docs covering system prompt logic, memory and model settings.
- Requires macOS 26 (Tahoe) on Apple Silicon, so Intel Macs and older systems are excluded.
- Not a standalone runtime: you must install the Pi CLI first, and Pi's behavior caps what you get.
- The install script pipes curl output into bash; you need to audit it yourself.
- Production builds send a launch event to EU-hosted PostHog, which telemetry-averse teams must weigh.
- Building from source needs Xcode 26.4+, and model usage costs are yours to pay.
- Documentation lives inside the repo under
agent-deck-documentation/and is awkward to browse outside it.
How does this agent compare with similar options?
The README explicitly positions Agent Deck against the Pi CLI: Pi is the coding agent runtime, while Agent Deck is a native macOS workspace around it that is not a replacement runtime and not a web wrapper. No other competitors are named.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Agent Deck This agent | 45 · Major gaps | Desktop appFree + model costs | ★ 124 | 3mo ago | Swift | — |
| Kooky | 68 · Some gaps | Desktop appFree | ★ 657 | 1mo ago | Swift | Codex · Claude Code |
| agterm | 75 · Some gaps | Desktop appFree | ★ 759 | today | Swift | Codex · Claude Code |
| zerostack | 69 · Some gaps | CLIFree + model costs | ★ 1.8k | 2d ago | Rust | OpenAI API · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
Evidence shows deliberate write-target design: built-in resources read-only, customizations via overrides, report-only subagents write only their own artifacts, and currentCheckout requires explicit confirmation (consistent across loop-authoring, AGENTS.md, README). However the MCP skill lets an agent directly rewrite ~/.pi/agent/mcp.json with a shell upsert template and may store secrets, so least privilege and sensitive-data handling are only partial. Dependency security shows only a pinned Sparkle version plus signing/notarization; no dependency audit or vulnerability mitigation is documented. Rollback relies on git worktrees and Sparkle updates with no explicit undo/recovery flow. Source attribution only credits the Pi upstream in README without finer provenance. Scores are deducted accordingly rather than awarded full marks.
README, AGENTS.md and skill docs are mutually consistent on Pi JSONL RPC, resource scopes and write targets, giving good self-consistency. Dependency availability rests only on Xcode 26.4+/macOS 26/Apple Silicon and SwiftPM resolution, with no lockfile or offline/degraded-path notes. Failure messages are explicit only in release.yml (Xcode selection, signature verification); runtime error handling is not evidenced.
Single scenario: macOS 26 Apple Silicon, clearly described. Capability boundaries are explicit (Pi is the runtime, Agent Deck is the workspace, not a replacement). Trigger precision relies only on skill descriptions and whenToUse conventions with no verifiable trigger rules. Environment fit is very narrow (macOS 26 + Xcode 26.4+ only) with no cross-platform or older-OS guidance.
Information architecture is clear (README + topic-split agent-deck-documentation + skills directory). Install notes cover both curl-pipe and DMG paths and warn to read the script first. License is MIT and complete. Naming stability is only a kebab-case convention with no versioned naming policy. Examples are screenshots and short snippets with no FAQ. Known limitations are scattered (bundled loop templates disabled, no by-reference agent import) rather than consolidated. Versioning is only release tags plus a Sparkle appcast with no CHANGELOG. Maintenance responsibility cannot be confirmed because the publisher is unverified.
Output usability is good: skills provide copyable frontmatter, JSON configs and upsert script templates. Marginal value is real: it organizes Pi sessions, resource scopes and parallel worktrees into a visible workspace beyond the bare CLI. Cost-benefit is limited by macOS 26 Apple Silicon only, requiring an installed Pi and Xcode to build, which raises the bar for ordinary users.
The README privacy statement (PostHog EU, app_opened only, disabled in Debug) and the release.yml signing/notarization/Sparkle injection steps corroborate each other, but both are self-reported with no independent tests or third-party audit. Fact/inference separation is only fair: docs mix product promises with implementation detail without marking what is verified. Cross-source corroboration is limited to in-repo files with no external verifiable source.
- The MCP install skill lets an agent directly rewrite ~/.pi/agent/mcp.json and may store secrets; review written content and secret handling manually.
- README recommends a curl-pipe install script; even with a read-first warning this is high-risk, prefer the signed DMG.
- Supports only macOS 26 Apple Silicon, a very narrow environment fit; older-OS or Intel users are excluded.
- Publisher identity is unverified, so maintenance responsibility and update path cannot be confirmed; rely on long-term maintenance with caution.
- No documented dependency security or vulnerability mitigation beyond a pinned Sparkle version.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
v5.1.3 Agent Deck 5.1.3
Release notes excerpt
What's new in 5.1.3 🐛 Bug fixes - Attachment pills now respond reliably when clicked.View on GitHub ↗ -
v5.1.2 Agent Deck 5.1.2
Release notes excerpt
What's new in 5.1.2 ✨ New features - Preview agent profiles from new hover cards. 💪 Improvements - Cleaner assignment controls with centered checkboxes and visible collection updates. 🐛 Bug fixes - Fixed the agent profile popover layout.View on GitHub ↗ -
v5.1.1 Agent Deck 5.1.1
Release notes excerpt
What's new in 5.1.1 🐛 Bug fixes - Unsent composer drafts now stay with their session when switching chats and remain available until the app closes. - Message cards now size styled text and attachment labels correctly without premature wrapping.View on GitHub ↗
FAQ
Does Agent Deck call an LLM itself?
Can I run it on an Intel Mac or on Windows?
Will files sitting in my project folder join the session automatically?
How do parallel tasks avoid stepping on each other?
Does it collect my prompts or code?
app_opened event per launch with no prompts, responses, code or file contents, project names or paths, account identifiers, credentials or screenshots; PostHog is EU-hosted.Is Agent Deck free?
Is Agent Deck open source?
Is Agent Deck safe to use?
Is Agent Deck still maintained?
Who makes Agent Deck?
What is Agent Deck's official website?
Thanks — we use this to improve the page.