Multi-AI Chat Desktop

English · 繁體中文 · 日本語 · Deutsch

Ask one question, then let your logged-in ChatGPT, Claude, Gemini, and Grok web sessions answer, review, challenge, and refine one another. Multi-AI Chat Desktop is a Tauri 2 workflow hub—not four chat windows placed side by side.

Current release: download the latest stable version · MIT · no API keys · no analytics

This project automates provider web pages you already use. Provider UI changes can temporarily break an adapter, and automated use may be subject to each provider’s terms. Use accounts and content you are authorized to use.

Project status: Feature development is complete. The final optional AI-Sister four-character commemorative theme and its 12-round Brainstorm preset are included; future changes are limited to provider compatibility, security, and build breakage. Brainstorm keeps four rotating seats and 48 contributions with full same-session history. Every built-in four-role or four-seat setup assigns ChatGPT, Claude, Gemini, and Grok once each. The shipped snapshot/replay tools remain available as-is with no further roadmap.

v1.8.4 highlights

• Complete provider answers. Finish-time DOM text is authoritative, code blocks containing replacement patterns remain intact, and an unfinished ChatGPT turn cannot silently become a successful partial response.
• Bounded Grok login recovery. Closing an allowed Grok authentication popup can trigger one native reload only while the same document remains blocked; lifecycle ownership and a start lease prevent duplicate or permanently wedged reloads.
• Provider-specific challenge guidance. Grok directs users to complete the solvable check in its pane, while Gemini keeps the required system-browser guidance for Google /sorry blocks.
• Correct portable update instructions. The portable README links directly to GitHub Releases because portable mode hides the in-app update controls.

See the bilingual v1.8.4 release notes for validation evidence, contributor credit, and the remaining manual release gates.

Choose the right edition

Edition	Best for	How it runs
Desktop (this repo)	Full workflows, focused live provider view, replay, snapshots, local files	Tauri app with isolated local provider profiles
Browser extension	Lightweight workflows inside Chrome	Chrome Side Panel controlling your existing provider tabs

What the desktop edition includes

• Reliable offscreen automation. Providers keep working without manually opening each “live page”; rejected sends retry and fail clearly instead of waiting forever.
• Conversation-first layout. Workflow controls sit above the less-important provider WebView on the left; the larger transcript pane can expand across the whole window, and provider chips identify the answer at the reading line.
• Six guided presets, five stable modes. Free distribution, debate, consultation, coding, five-round truth-seeking roundtable, and a 12-round × 4-seat Brainstorm workflow with 48 contributions.
• Configurable collaboration roles. Built-in four-role setups use all four providers once each. In Settings, you can assign roles differently; serial roles may reuse one provider, while roles that run in parallel must remain distinct.
• Local sessions. Create a new conversation or reopen up to 30 recent transcripts stored on this computer; restored follow-ups receive bounded context from the same session.
• Readable results. Safe semantic Markdown rendering for headings, nested lists, links, quotes, fenced code, and scrollable tables.
• Image completion. Image-only ChatGPT responses complete the workflow instead of hanging.
• Reproducible work. Optional snapshots, privacy tiers, replay, provider diagnostics, and a 2,000-event deduplicated log.
• Four UI languages. English, Traditional Chinese, Japanese, and German.
• Question-aware response language. Auto follows an explicit request, then the question and conversation language; the UI language is used only as a fallback. A fixed response language can be selected separately.
• AI-Sister Commemorative Edition. One optional four-character theme adds the supplied portraits to provider cards, active speakers, process rows, and the app shell without reskinning third-party pages.
• Repository Skills. Codex and Claude Code can validate prerequisites and launch the source app without an installer.
• Safer macOS packaging. Apple Silicon DMGs are ad-hoc signed, and release CI verifies the embedded app signature before upload.

Workflow presets

Preset	Flow	Typical use
Free	Selected providers answer in parallel	Quick comparison and image prompts
Debate	Pro → Con → Judge → Synthesis	Stress-test a decision or argument
Consult	Two independent answers → Review → Final answer	Research and second opinions
Coding	Specification → Reviews → v1 → Tests → v2 → Acceptance → Final	Structured software planning and review
Roundtable	5 rounds × 4 seats = 20 turns; all 4 providers by default	Slow, adversarial convergence on difficult questions
Brainstorm	12 rounds × 4 rotating seats = 48 contributions; framing → divergence → cross-pollination → harvesting → concept tests	Full-context idea development, a balanced portfolio, and concrete first experiments

Structured workflows retry a provider once when its page rejects a just-finished back-to-back send. Any remaining engine error stops the workflow instead of being passed to later roles as if it were an answer.

Brainstorm is intentionally the heaviest preset: keep all four default provider sessions authenticated and allow roughly 45–90 minutes. If a provider is unavailable, preflight identifies it and you can customize the role assignments. Claude's current consumer site requires an account login; this app detects and guides that flow but never bypasses provider login or security checks.

After a workflow finishes, use the bottom composer to continue the same conversation. Use New conversation when you want a clean session.

Install a release

Download from Releases:

• Windows x64: portable .zip or x64-setup.exe. Windows 10/11 normally already includes WebView2; the installer can fetch it when missing.
• macOS Apple Silicon: aarch64.dmg. Builds from v1.0.1 onward are ad-hoc signed but not Apple-notarized. Intel builds are not currently published.
• Linux x64: .AppImage, then run chmod +x Multi-AI*.AppImage and open it. Ubuntu 22.04 / Debian 12 or newer is recommended.

On first launch, open each provider once and sign in. Credentials stay in that provider’s local WebView profile; the app never asks for the password.

macOS first launch

1. Delete any v1.0.0 copy, download v1.0.1 or newer, open the DMG, and drag the app to Applications.
2. Try to open the app once.
3. Within about one hour, open System Settings → Privacy & Security, scroll to Security, then choose Open Anyway and confirm.

The ad-hoc signature prevents the false “app is damaged” bundle-integrity failure, but only Apple Developer ID signing plus notarization can remove the first-launch security exception entirely. Managed Macs may prohibit user exceptions.

Launch the source with Codex or Claude Code

The repo contains two explicit local Skills:

• Codex: .agents/skills/launch-multi-ai-chat/SKILL.md
• Claude Code: .claude/skills/launch-multi-ai-chat/SKILL.md

These follow the official Codex Agent Skills and Claude Code Skills layouts.

The machine-readable source of truth is agent-release.json, validated by agent-release.schema.json. The full trust, permission, side-effect, readiness, and audit model is documented in the bilingual Agent-Ready Source Release contract.

Opening a repository never executes it automatically. Source launch executes this checkout, JavaScript dependency lifecycle code, and Rust build scripts/procedural macros, so review and trust the repo first. The explicit Skill can install locked dependencies into this project, build generated code, and start tauri dev; it never installs or removes host toolchains/global packages, changes PATH or security settings, builds an installer, or reads provider credentials. Any host installation is a separate operation requiring separate explicit approval.

Codex app, CLI, or IDE

1. Download/clone this repository and open the folder as a local Codex project/task.
2. Type $launch-multi-ai-chat (or choose Launch Multi-AI Chat from /skills).
3. Approve local command execution if your Codex security settings request it.
4. Wait for the first Rust build; the Tauri window opens when compilation finishes.

Repo Skills work in Codex app, CLI, and IDE surfaces. A remote/cloud task can edit this repo but cannot display a GUI on your computer.

Claude Code desktop, CLI, or IDE

1. Open this repository in a Claude Code surface that has a local shell on your graphical computer.
2. Run /launch-multi-ai-chat.
3. Keep the repository folder available while the dev app runs.

If your Claude desktop/browser session is remote, use a local Claude Code session or run claude from a terminal in this folder, then invoke /launch-multi-ai-chat.

Platform prerequisites for source launch

Common: Node.js ^22.13.0 || >=24.0.0, pnpm (or Corepack), and the stable Rust toolchain. The commands below are manual prerequisite examples; the Skill only reports missing items and stops.

Windows 10/11

1. Install Node.js LTS.
2. Install Rust with winget install --id Rustlang.Rustup and select the MSVC toolchain.
3. Install Visual Studio Build Tools → Desktop development with C++.
4. Install Microsoft Edge WebView2 Evergreen Runtime only if it is missing.

macOS 10.15+

1. Run xcode-select --install (full Xcode is not required for desktop-only development).
2. Install Node.js LTS and Rust stable.
3. Start the Skill from a local graphical login session, not SSH.

Ubuntu / Debian

sudo apt update
sudo apt install libwebkit2gtk-4.1-dev build-essential curl wget file \
  libxdo-dev libssl-dev libayatana-appindicator3-dev librsvg2-dev

Then install Node.js ^22.13.0 || >=24.0.0, Rust stable, and run the Skill from a graphical X11/Wayland session. Other distributions should follow the Tauri 2 prerequisites.

Skill lifecycle commands

node scripts/agent/audit.mjs --phase before --write --json
node scripts/agent/doctor.mjs --json
node scripts/agent/launch.mjs --dry-run --json
node scripts/agent/launch.mjs --wait --timeout-ms 600000 --json
node scripts/agent/status.mjs --json --lines 80
node scripts/agent/audit.mjs --phase after --write --json
node scripts/agent/stop.mjs --json
pnpm agent:verify

The first Rust build can take several minutes. accepted and building are not readiness claims: only the current run's [MAC_AGENT] READY control-pane marker produces state: "ready". Logs, identity state, and before/after receipts stay under ignored .agent-runtime/; nothing is uploaded automatically. This GUI/WebView lane intentionally has no Docker variant.

Development

corepack enable        # only if pnpm is not already available
pnpm install --frozen-lockfile
pnpm build:injected
pnpm verify
pnpm tauri dev

pnpm tauri build creates platform packages. See docs/SPEC.md, docs/ARCHITECTURE.md, docs/RELEASE.md, and the honest compatibility matrix.

Privacy and network behavior

• No API keys, project account, telemetry, or conversation backend.
• Prompts go directly to the provider pages selected by the user.
• Provider cookies and profiles stay in local app data.
• Adapter updates are optional, data-only, schema-validated, and unable to expand the URL scopes bundled with the app.
• Debug bundles are created locally only when requested.
• Export/share actions run only after an explicit user action.

Project

Report vulnerabilities privately through SECURITY.md. Report provider automation regressions with the GitHub Adapter broken issue form after reviewing the in-app diagnostic preview.

Contributors and acknowledgements

Special thanks to Dave Tseng (@DaveTseng2019) for the v1.3.1 overlay reliability fix, the careful reproductions and original proposals in #10, #11, and #12, the serializer regression tests merged in #14, the Grok challenge and expanded-focus work in #39 and #40, and the full-width transcript plus scroll-linked provider focus contributed through #51.

Thank you to CE Lin (@ChingEnLin) for the detailed provider-status report in #41 and the ChatGPT, Gemini, and Grok adapter repair contributed through #42.

Thank you to the Windows and macOS users who shared reproducible reports and sanitized debug logs. Those reports directly improved first-launch packaging, provider automation, session continuity, and release verification.

Sponsored by AI-Sister.com. Created by Ted Huang (TED@TED-H.com, ted-h.com).

MIT License.