Pairlet

Pairlet was formerly called CC Pocket (cc-pocket). This is the same project, with its history and existing download links preserved.

English | 简体中文

Your coding agents stay on your computer. You stay in control from anywhere.

Pairlet is an open-source, local-first control plane for command-line coding agents. The agent keeps running on your own machine, against your own checkout; from a phone, a tablet or another computer you watch it work, answer the permission prompts that block it, continue the same session, and read what it changed. Traffic is end-to-end encrypted and passes through a zero-knowledge relay that only ever forwards ciphertext — no Pairlet account, no content logging. Clean-room Kotlin, MIT.

v1.9.4 drives six agent backends — Claude Code, OpenAI Codex, OpenCode, Kimi Code (Preview), ZCode and DeepSeek. They are not equivalent: see the capability matrix before you pick one.

🌐 Website · 📖 User manual · 💬 Support, no sign-in · 📦 Latest release

Real product UI with scripted demo data — regenerate with bash marketing/site/generate-assets.sh. Provenance: site/assets/product/manifest.json.

Development builds now include Firebase/GA4 product outcomes and Sentry diagnostics, with controls for sharing usage and diagnostic data. See the observability guide for configuration, privacy boundaries, verified results and remaining release requirements.

Quick start

1 · Get the app — App Store (iPhone · iPad) · TestFlight beta · Android APK. Prefer a computer? See the Desktop app.

2 · Install the daemon on the machine that runs your agent CLI — any supported one, not Claude specifically:

curl -fsSL https://raw.githubusercontent.com/heypandax/pairlet/main/scripts/install.sh | bash   # macOS · Linux
irm https://raw.githubusercontent.com/heypandax/pairlet/main/scripts/install.ps1 | iex          # Windows

3 · Pair — run pairlet pair, then scan the QR it prints (or type the 6-digit code) in the app. You are connected end-to-end.

pairlet is the short command; cc-pocket-daemon still supports the same subcommands. If an older installation has no pairlet command yet, keep using cc-pocket-daemon until you update through its original installer or package manager.

Package managers, mirrors, updates and per-platform notes: Install details.

The four jobs

Capability differs by backend — the matrix below is the source of truth.

Agent support

Public capability claims for v1.9.4, audited against commit e9ee816f on main. Machine-readable copy: site/public-capabilities.json.

• Core session means discover, replay, create, resume, send and receive text, and live streaming. Every backend does all of it.
• OpenCode has no enforceable interactive approval. opencode run has no approval protocol, so those sessions run at Full access and the app says so up front instead of offering modes it cannot enforce.
• Kimi Code is Preview. DeepSeek is supported, but narrow: approvals and multiple-choice questions are bridged to the app, but the sandbox mode is fixed at launch (changing it relaunches the session), and there is no Changed-files/diff view, no usage accounting and no model switching.
• DeepSeek Harness has no timeout of its own. Left alone, an unanswered approval or question blocks its turn indefinitely — it does not deny. Pairlet puts the request on the daemon's normal approval window instead: an approval that expires is answered reject, and a question that expires is answered skipped, so an unanswered request ends the wait rather than hanging it. DeepSeek also has no "always allow" — every request is a one-off decision.
• Boundaries follow the release. Full detail: Features and the User manual.

Architecture & trust boundary

flowchart LR
    app["📱🖥️ Pairlet app<br/>(phone · tablet · desktop)"] -- "wss · ciphertext" --> relay["relay<br/>(zero-knowledge broker)"]
    relay -- "wss · ciphertext" --> daemon["local daemon<br/>(your computer)"]
    daemon -- "stdio" --> agent["agent CLI<br/>(claude · codex · opencode · kimi · zcode · DeepSeek)"]

The daemon runs on your computer, drives the agent CLI as a subprocess and dials out to the relay — no inbound ports to open. The relay pairs your devices and routes opaque encrypted frames; it holds no message content and no private keys. The app and the daemon run an end-to-end session (P-256 ECDH + HKDF + AES-256-GCM, an X3DH/Noise-style handshake), so plaintext never leaves the two trusted endpoints. On the same network the app connects to the daemon directly for lower latency; the relay stays as the from-anywhere fallback. Pairings expire and can be revoked.

Honest limits: the agent still executes with your own operating-system permissions — end-to-end encryption is not a sandbox. OpenCode sessions have no enforceable interactive approval. The custom Noise-style channel has not had an independent third-party audit. Threat model: docs/SECURITY.md. Report vulnerabilities privately via GitHub security advisories.

Platforms & distribution

The desktop app and the local daemon are different packages: the app is a client, the daemon is what actually runs the agent.

Install details

curl -fsSL https://raw.githubusercontent.com/heypandax/pairlet/main/scripts/install.sh | bash
pairlet pair

Verifies the download against the release's SHA256SUMS, installs under ~/.local (one directory per version), and registers the launchd service so it runs on login and reconnects itself. Homebrew: brew install --cask heypandax/tap/cc-pocket (use the full name; an unrelated cask is also called cc-pocket).

curl -fsSL https://raw.githubusercontent.com/heypandax/pairlet/main/scripts/install.sh | bash
pairlet pair

Pulls a self-contained tarball (bundled JRE, no system Java), installs under ~/.local and registers a systemd --user service. Voice transcription uses ffmpeg instead of macOS's afconvert. There is no official Linux desktop app package — build it from source.

<det