Rish, your pocket agent.

Run locally. Choose your model.
_{Local workspaces · Model choice · Tool execution · Approvals}

DSH · Claude Code · Codex · GLM
_{Built-in connections · Native Rish adapters}

English · 简体中文 · 繁體中文 · 日本語 · 한국어 · Français · Español · Deutsch · Português · Русский · हिन्दी · Türkçe · ไทย · Tiếng Việt · Bahasa Indonesia

Get started · Built-in connections · Product tour · Platforms and models · Ecosystem · Developer guide · MIT License

Rish brings Agent conversations, workspaces, and tool execution to your phone. Choose a model, describe a task, inspect the work, and approve changes without keeping a computer running. Coding is one of its uses, not its only purpose.

Preparing an experimental source preview; no stable installable release yet. Platform scope and account/subscription verification status are summarized in Platforms and models.

Built-in connections

Four built-in connections. One pocket workspace.

DSH DeepSeek · Editable model catalog

Claude Code Anthropic · API key / Subscription sign-in¹

Codex OpenAI · API key / Subscription sign-in¹

GLM Zhipu · API key / Subscription sign-in¹

¹ Subscription sign-in is currently iOS-only. BigModel Coding Lite is verified; Codex and Claude Code require the optional experimental build. See the verification details below.

Choose a harness and connect with an API key or an account supported by your build. Then start working with files and projects on your phone. Rish manages the agent loop, workspace, tool approvals, and execution records; the built-in adapters connect to model services.

Account and subscription verification (iOS)

• Codex: the optional experimental build has verified official-CLI device login, a subscription gpt-5.6-luna text chat, a local list_dir tool call, and persistence across restart. The official CLI is used only for login; not all tools and models are verified.
• GLM: ZCode is Zhipu's agent product; GLM is the model family. BigModel sign-in, restart persistence, and a Coding Lite GLM-5.3 response are verified; the trial allowance is not. The official ZCode runtime is not integrated.
• Claude Code: the optional iOS experimental build has verified subscription login, a Haiku 4.5 text response through the unmodified official CLI, and persistence across restart. This path currently supports text only, without tools or attachments. A measured turn took about 4.5 minutes; performance still needs work.

Product tour

Follow Agent execution, review project changes, and choose a model connection. Click a screenshot to open the original.

Agent conversation — Follow progress, tool calls, and results on the phone. A path outside the workspace is refused, and the model corrects itself in the next round.	Local projects — Inspect unstaged files and change statistics before committing.
iPad workspace and model connections — A wide-screen sidebar, dark appearance, and API adapter entries.

All captures are from actual Simulators, showing the current UI and workflow.

Why Rish

Put it to work

These examples use the currently available iOS capabilities. Supported tools and file formats vary by platform. For the Linux Guest experiments, see the runtime reference.

How local execution works

Your task → Rish assembles context → Your chosen model service
                                           ↓ Text / tool requests
Phone workspace ← Local tools ← Rish validation and approval

Supported tools execute in the phone's app-owned environment. Model requests send selected conversation and task context to your configured service: local execution does not mean offline model inference.

Rish combines native file/Git operations, the Rish runtime, and an experimental Linux Guest. Full desktop-program compatibility and indefinite background execution are not promised. The developer guide separates verified capabilities from experimental paths.

Platforms and models

Ecosystem

Rish is part of a family of local-first, AI-native tools from ZSeven-W. rish boots the Linux guest inside this app; the others carry the same idea onto other surfaces — the terminal, the design canvas, and an agent's memory.

Project	What it is
rish	Real Docker on a phone: a no-JIT x86-64 full-system interpreter in pure Rust that boots Linux and runs containers on iOS and Android. This app's Linux Guest comes from here.
OpenPencil	The first open-source AI-native vector design tool, and the first with concurrent Agent Teams. Design-as-Code — turn prompts into UI directly on the live canvas.
jian	Pure-Rust, GPU-Skia UI framework. Turns a declarative .op document into a native app — no JS runtime, no DOM, no Electron.
Zode	AI-native coding CLI for your terminal. A fast Rust TUI that reads your code, runs commands, searches files, and manages git.
noema	Local-first, non-vector memory for coding agents. Durable memory as inspectable files, a review queue, and embedding-free recall.

Get started

Build from source for now; there is no stable end-user download. Once you have the source, start at the repository root:

node scripts/verify-source-checkout.mjs
npm ci --prefix apps/mobile

iOS: Native preparation requires pinned Xcode, Rust, and SDK versions. Read the prerequisites, then run:

./scripts/prepare-rish-ios.sh
./scripts/prepare-libgit2-ios.sh
cd apps/mobile/ios
pod install
cd ../../..
npm run ios --prefix apps/mobile

Android: With an Android development environment configured, run npm run android --prefix apps/mobile. The developer guide also covers standalone test APKs.

Open the app, choose a model, and connect with an API key or a supported account. On iOS, create or select a project, review its context, and start a task. Codex and Claude Code subscription sign-in requires an optional experimental build (see the developer guide); see the account guide for BigModel. Credentials stay in native secure storage.

Progress and contributing

The first source preview is in preparation. Future releases will be marked Pre-release. Complete Harness compatibility, Android local execution, and continuous background operation remain limited. See the preview scope and roadmap.

Contributions to documentation, platform support, model compatibility, and reproducible fixes are welcome. Read CONTRIBUTING first. For security issues, consult SECURITY before sharing details; never post credentials or sensitive data publicly.

• Developer and build guide
• Brand and approved copy
• Third-party notices and Guest sources

Project code is licensed under MIT. Third-party runtimes, Guest components, and other dependencies retain their own licenses.