Superagent
A persistent Mac workspace that puts your coding agent beside a controllable browser, project files, and an iOS Simulator.
- Source repo
- pungme/superagent-desktop
- Stars
- ★ 26
- Last updated
- 1d ago
- License
- MIT
- Primary language
- TypeScript
- FA score
- 58/100 · Major gaps
At a glance
- Works with
- Portable with changesCodex · Claude Code
- You'll need
- Typical use
- A Mac-based web developer using Claude Code or Codex who wants to watch code changes and the resulting local page update in the same window.
- Main limitation
- The desktop application explicitly requires an Apple Silicon Mac; no Intel Mac, Windows, or Linux build is documented.
- Source review
- 58/100 · Major gaps
What does this agent do, and when should you use it?
Superagent is an Electron, TypeScript, and React desktop application for Apple Silicon Macs that gives Claude Code or Codex a persistent visual workspace, with the agent selectable per chat. A project conversation can sit beside a real browser, the file tree, and an optional streamed iOS Simulator, allowing the agent to modify software and inspect the running result in one window. Each new conversation can receive a private project checkout and branch on its first message; Keep folds the result back as one change, while Throw away removes it. Conversations, panes, open files, and active state survive restarts, while scheduled Routines, project boards, PDF annotation, local speech transcription, and context usage are built into the desktop interface. Core execution stays on the Mac and uses the user's existing agent subscription without a separate API key; optional phone access connects the companion iOS app through an end-to-end encrypted blind relay.
After the user submits a task in a project chat, Superagent starts the selected coding agent and lets it read and edit project files, run commands, start a local development server, and update code on the conversation's isolated branch. The agent can operate Superagent's own browser pane by opening pages, clicking, typing, and reading results; automation does not enter the user's personal browser. For iOS work, tools including sim_list_devices, sim_boot, sim_screenshot, sim_open_url, and sim_install_and_launch call xcrun simctl, while an optional stream displays the Simulator inside the application; this workflow requires Xcode. A completed checkout can be folded into the project as one change with Keep or deleted with Throw away. Routines repeat browser checks on natural-language schedules, and the file viewer opens source, Markdown, images, and PDFs, including in-place PDF annotation. Once a phone is paired, the Mac and iPhone exchange conversations, pages, prompts, and approvals through superagent-relay using end-to-end encryption with a per-device key; the relay is described as unable to read or store the plaintext.
- A Mac-based web developer using Claude Code or Codex who wants to watch code changes and the resulting local page update in the same window.
- A developer who needs an agent to click, type, and read within authenticated websites while keeping automation isolated from their personal browser.
- An iPhone application engineer with Xcode installed who wants the agent to build, launch, inspect, and iterate against an iOS Simulator.
- A developer exploring several features or fixes concurrently who wants each conversation isolated in its own checkout and branch.
- A user who needs to answer agent questions or approve actions away from the desk through the companion iPhone or iPad application.
- An individual developer who wants scheduled browser checks expressed as natural-language Routines, such as checking a site every hour.
How do you install or deploy this agent?
The prebuilt application requires an Apple Silicon Mac. Download SuperAgent.dmg from the releases page, open the disk image, and install the signed and notarized application. For Claude Code chats, install Claude Code and sign in first; the product also explicitly advertises per-chat Codex support, although the supplied material does not give a separate Codex installation command. Install Xcode only if you need the iOS Simulator. For remote phone access, install the companion application through TestFlight or build pungme/superagent-ios from source, then pair it under Settings → Phone. To run the desktop application from source, execute cd app, npm install, and npm run dev. Run unit tests with npm test; npm run test:e2e builds the application and drives it with Playwright.
How do you use this agent?
Add a project, create a conversation, and describe the coding task in the first message; Superagent then creates that conversation's private project copy and branch. Select Claude Code or Codex for the chat and choose the corresponding model and permission mode. Ask mode requests approval before the agent acts, either on the Mac or on a paired phone. Point the browser pane at a local development server or live site; the agent can start the server and interact with the visible page. For iOS projects, ask it to build and launch in Simulator. Use the scissors control on the browser or Simulator, or press ⌘⇧S, to crop a region into the message. When the work is ready, choose Keep to fold it back as one change, or Throw away to delete the conversation's working copy. Pair a phone under Settings → Phone for remote interaction, and describe a timed browser task in plain language to create a Routine.
What are this agent's strengths and limitations?
- The browser is visible and user-takeover-friendly rather than a hidden automation instance, while still preserving its own authenticated session.
- Per-conversation checkouts and branches isolate concurrent experiments, with explicit Keep and Throw away outcomes.
- Coding chat, live diffs, files, browser content, PDFs, and an iOS Simulator share one restorable desktop workspace.
- Local execution reuses an existing Claude Code or Codex subscription and does not require a separate model API key.
- Phone communication uses per-device end-to-end encryption, and users may run their own
superagent-relay.
- The desktop application explicitly requires an Apple Silicon Mac; no Intel Mac, Windows, or Linux build is documented.
- Simulator support requires Xcode, and live framebuffer capture relies on a native helper and private framework path with a slower screenshot fallback if it stops working.
- The browser can act within real authenticated sessions. Although automation is confined to Superagent's pane, adopters must still assess account permissions and automation risk.
- Remote phone access adds a companion application and relay connection; installations that do not pair a phone do not receive that capability.
- The supplied documentation is inconsistent about setup scope: the product text says Claude Code and Codex are switchable per chat, while the requirements section only details installing and signing in to Claude Code.
- Adapters for additional CLIs and local models remain roadmap items, so the current application is not a general-purpose agent host.
How does this agent compare with similar options?
Compared with running Claude Code or Codex only in a terminal, Superagent adds persistent desktop state, a visible and controllable authenticated browser, per-chat checkouts, an embedded iOS Simulator view, and remote phone interaction. It still runs those existing coding agents on their existing subscriptions and does not supply an AI model of its own. Support for other CLIs and local models is described as future work, so the present release should not be treated as a universal desktop shell for arbitrary agents.
Key facts side by side with the most closely related agents.
| Agent | Source review | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|
| Superagent This agent | 58 · Major gaps | ★ 26 | 1d ago | TypeScript | Codex · Claude Code |
| Nimbalyst Visual Coding Workspace | 50 · Major gaps | ★ 1.8k | 1d ago | TypeScript | Codex · Claude Code |
| Munder Difflin – Multi-Agent Office Floor | 43 · Major gaps | ★ 7.9k | 1d ago | TypeScript | ChatGPT · Codex · Claude Code |
| Apache Maka | 82 · Good | ★ 5.6k | today | TypeScript | — |
How does FollowAgents rate this agent?
Why each dimension lost points
The README gives reasonably specific accounts of network destinations, conditional phone pairing, browser isolation, local transcription, per-chat checkouts, and discard behavior. Maintainer notes additionally describe pinned dependency hashes, signing, notarization, and updater integrity checks, supporting partial credit for least privilege, dependency security, data-flow disclosure, and rollback. Deductions apply because these are mostly assertions rather than implementation evidence in the supplied files. Ask mode is optional, with no demonstrated mandatory confirmation for high-risk browser, filesystem, scheduled, or remote actions. The @~/ feature exposes a broad disk scope, while authenticated browsing and routines can create external effects. End-to-end encryption, blind-relay behavior, non-retention, and sensitive-data lifecycle controls are not verified by supplied code. Licensing and related repositories are attributed, but the responsible maintainer identity remains unclear.
Dependencies and fallback paths are described usefully: Apple Silicon, Claude Code, optional Xcode, bundled baguette with a brew fallback, and screenshot fallback when simulator streaming fails. The principal deduction is internal inconsistency. The introduction advertises both Claude Code and Codex, the requirements name only Claude Code, and the roadmap says the product currently wraps Claude. The product text says the agent can operate the simulator, while the maintainer note says sim_tap, sim_type, and sim_swipe remain unfinished. User-facing failure messages and recovery behavior receive only scattered treatment.
The audience and scenarios are covered thoroughly across coding, authenticated browsing, iOS development, mobile participation, scheduled routines, file viewing, and multiple projects, justifying full credit there. Platform, network, and optional-component boundaries are mostly explicit, but Codex support and simulator-control boundaries conflict. Natural-language routines and broad browser/filesystem reach are not accompanied by evidence of scope resolution, scheduling-conflict handling, or false-trigger prevention, reducing trigger precision. Environment fit is clearly centered on Apple Silicon macOS, but supported macOS, Node/npm, and agent CLI versions are not specified.
The README has strong information architecture, organizing features, requirements, network activity, related repositories, source setup, roadmap, and licensing, with visual examples. Basic development and test commands are present, but Node/npm versions, complete build instructions, and troubleshooting are absent. Capitalization varies among Superagent, SuperAgent, and artifact names, and no dedicated FAQ is supplied. Limitations and future work are disclosed, especially in detailed notes about private APIs, streaming failures, and release hazards, but important caveats are hidden in non-user-facing comments. The complete MIT license earns full credit. Release badges and a stated in-app What's New mechanism do not substitute for a supplied changelog. A contribution link exists, but 'Superagent contributors' does not establish clear maintenance ownership or support responsibility.
Persistent conversations, isolated checkouts, an observable browser, simulator integration, mobile participation, and local dashboards form a plausibly useful unified workflow, with work intended to be inspectable, kept, or discarded. Scores are reduced because nearly all effectiveness evidence is descriptive rather than backed by supplied implementation or output examples, and conflicting Codex and simulator-control claims weaken expected usability. Reuse of an existing agent subscription limits incremental financial cost, but Apple Silicon exclusivity, optional Xcode and iPhone requirements, external CLI dependencies, and supervision of broad automation add meaningful cost.
Roadmap items are generally distinguished from shipped claims, and maintainer notes record versions, dates, failed experiments, corrected conclusions, and release incidents, which supports fact-versus-inference separation. Most security and capability claims, however, occur only in the README. The sole supplied code file is a localStorage test shim and cannot trace browser control, encryption, network restrictions, permissions, rollback, or updating to implementation. The LICENSE corroborates licensing and the shim modestly corroborates test infrastructure, but cross-source support for the central product claims is otherwise thin.
- Do not treat end-to-end encryption, relay non-retention, or the claimed exhaustive outbound-network list as proven guarantees without reviewing their implementations.
- Authenticated browser control, @~/ disk reach, and scheduled routines create a broad external-effects surface. Verify per-action authorization, stopping, auditability, and irreversible-action safeguards before using sensitive accounts or data.
- The README conflicts on current Codex support and on whether the agent can actually tap, type, and swipe in the simulator; confirm the revision's real capabilities before adoption.
- Simulator streaming uses private Apple APIs with a screenshot fallback, so macOS or Xcode changes may degrade or break the feature.
- Automatic updates affect installed users, and the maintainer notes document prior notarization, hash, and latest-release mistakes. Independently verify signing, notarization, and update metadata.
FAQ
Does it require another AI subscription or API key?
Can it control my everyday browser?
Can I use it without an iPhone or Xcode?
Does phone traffic pass through a third-party service?
What happens if native Simulator capture breaks?
simctl io screenshot.