PBJ-2/clauderipple

English · 한국어 · 简体中文 · 日本語 · Español · Português (BR)

Chat stays Claude. Only Code becomes GPT.
Claude Desktop's own setting for other models switches the whole app — chat, phone Remote Control and connectors go with it.
ClaudeRipple turns nothing off: you stay signed in to your Claude subscription while GPT, DeepSeek, Kimi, Grok
or 400+ other models answer in the Code tab. Claude inside the Codex app and Codex CLI too.

_{A one-minute tour: other models in Claude Desktop's own model menu, effort set for every prompt, and each subagent sent to the model that fits. Click for the full video.}

GPT, DeepSeek, GLM and Gemini in Claude Desktop's own model menu	GPT-6 Astra working in the Code tab, every request logged on the right
Subagents on DeepSeek and GLM, named in the background-task panel
The ripple mod: effort, context, cache timer, and which model answered with its cache hit
Warnings when the cache goes cold or a limit is near, and a handoff to a fresh chat	Quick settings: automatic effort, the judge, worker modes

Claude Desktop already ships an official way to use other models — the inference gateway setting, what the app calls third-party or 3P mode. Turning it on is not a per-session choice. It switches the whole app into a different deployment mode at launch, and that mode is a different product:

• The window stops loading claude.ai and loads a local bundle instead. claude.ai /api/ and /v1/ calls answer custom_3p_not_available 503.
• "Chat" is no longer claude.ai chat. It is a Claude Code local-agent session.
• Remote Control and side sessions are switched off outright (shouldEnableSessionsBridge() returns false).
• Anthropic's own claude.ai connectors, Claude Design, mobile continuity and chat search go with it.

There is no middle setting. "Chat on Claude, Code on the gateway" is impossible for anyone, because 1P mode routes nothing through the gateway at all. (All of this is read out of the app's own bundle; see docs/ARCHITECTURE.md §2.)

ClaudeRipple never touches that setting. The app stays signed in to your Claude subscription in 1P mode, and another model answers in the Code tab and its subagents — under its real name in the model picker, at the reasoning effort you chose. You do not trade your Claude account for a GPT one. You keep both, in one app.

Three things follow from staying in 1P, and no gateway-based tool can offer any of them:

• Code from your phone, on GPT. Remote Control is off in 3P. It is on here, so you can pick a model and send a task from a phone and watch the commits land.
• Connectors alive while another model works. 3P stubs claude.ai calls, which takes Anthropic's own connectors with them. Read a Google Drive doc and have GPT write the code.
• Cloud and side sessions. Gone in 3P. Untouched here.

If you pay for Claude, 3P is not an option — it throws away the half of the subscription you are paying for. This is the only way to keep it and still choose the model.

Yes — in the Code tab and its subagents. Install ClaudeRipple, sign in with a ChatGPT Plus or Pro subscription, and GPT (GPT-6 Sol, Luna, Astra, GPT-5.6 Terra and the rest) shows up under its real name in the model picker, with the reasoning effort you pick. The app itself is not modified and not switched into any other mode. The general chat tab stays Claude: no tool can route it without turning the whole app into gateway mode (see 1P, not 3P).

The same install covers the terminal claude CLI. ClaudeRipple is a local proxy that Claude Code is pointed at through two lines in ~/.claude/settings.json, so your skills, hooks, MCP servers, CLAUDE.md and subagents keep working while another model answers.

No — and you should not. That setting switches the whole app into 3P mode, which stops loading claude.ai, replaces chat with a local-agent session and turns Remote Control off. ClaudeRipple never touches it.

No. The app stays signed in to your Claude subscription, so chat, phone Remote Control, cloud sessions and Anthropic's own connectors keep working. That is the whole point of staying in 1P.

Yes. Model mapping is per model name: leave a name mapped to Claude and it answers as Claude. Most people map one or two names to GPT and leave the rest alone.

No. Any ChatGPT plan that includes Codex works — Plus, Pro and Business, and Enterprise or Edu where the workspace admin has turned Codex on. You sign in from the app and no API key is involved. The plan's usage limits apply; once they are used up, Codex credits you have bought carry on. API keys are for the other providers (DeepSeek, Kimi, GLM, OpenRouter and anything else with an Anthropic- or OpenAI-compatible endpoint).

GPT through a ChatGPT subscription, and DeepSeek, Kimi, GLM, Grok, Qwen and 400+ more through OpenRouter or a direct API key. Claude models keep working as they are.

Yes, on Windows and macOS, arm64 and x64. Install, dashboard, restart and uninstall are measured on Windows 11 arm64; the x64 runtime is measured under emulation. Linux works too where there is a systemd user session: install, a Code-tab request through the router, a restart in the middle of a streaming answer and uninstall are measured on Ubuntu 24.04 x64 with Claude Desktop 2.2553.13, and so is picker mode: models from cli.extraModels show in the Code-tab picker next to Claude's own.

No. The proxy runs on your own machine. Requests go to the provider you configured and nowhere else, and credentials stay in your home directory. See Privacy.

What ClaudeRipple does with Claude is small and checkable: a request that stays on Claude leaves your machine byte for byte as Claude Code sent it, and a request routed to GPT or another provider never reaches Anthropic at all. It works through the proxy and certificate settings Claude Code documents for corporate networks (HTTPS_PROXY, NODE_EXTRA_CA_CERTS — Enterprise network configuration), not by patching the app. The reports of banned accounts are about the opposite direction — a Claude subscription login used from another program. ClaudeRipple has one optional feature of that kind: Claude inside Codex using your subscription login. That use is subject to Anthropic's terms; an Anthropic API key avoids the question. Rotating several Claude subscriptions is likewise your call under those terms. Only Anthropic can say how an account is treated, so nothing here is a promise.

An AI Studio API key is the way Google supports, and the safe one. The Google account sign-in (Antigravity) is different: Google's Antigravity terms (section 6) and the Gemini CLI FAQ name using those credentials from another program as a breach. In February 2026 Google suspended accounts used this way — the suspension also cut off Gemini CLI and Code Assist — and said a second violation is permanent. ClaudeRipple offers the sign-in because others do and people asked, shows this warning before it starts, and does nothing to hide the traffic. Use a secondary account if you use it at all.

Yes. Each clauderipple login (or + Add ChatGPT account in the dashboard) adds one. When an account hits its usage limit, the same request moves to the next account and the spent one rests until its window resets; a conversation stays on the account that answered, to keep its prompt cache. This applies to the Claude Desktop Code tab, Claude Code, and GPT in the Codex app and CLI.

Yes. clauderipple codex on adds ClaudeRipple as a provider, and Claude models appear in the Codex app's and CLI's model list under their own names. They answer through your Claude Code login or an Anthropic API key. See Codex app and Codex CLI.

Yes, with a ChatGPT subscription and no API key: clauderipple image "a red paper boat" -o boat.png --aspect square. --ref FILE adds a reference image and --transparent a transparent background. One image takes about 30 seconds, and the subscription chooses the resolution and quality. Subagents ClaudeRipple generates are told the same call, so they can make images too. See Claude Code.

Claude Code's system prompt and tool descriptions are written for Claude, so another model reads instructions tuned for someone else, and it may behave a little differently than in its own client. ClaudeRipple tells each model its own name and effort, reshapes tool schemas a vendor would reject, and lets you append instructions per provider. A common setup is Claude as the main model with GPT or DeepSeek as subagents.

See the comparison table. In short: ClaudeRipple reaches the Claude Desktop Code tab while the app stays signed in to Claude, and shows other models in its picker under their real names. CC Switch also reaches the desktop app, but by turning on its gateway (3P) mode.

Yes. ClaudeRipple is open source under GPL-3.0. You pay only for the subscriptions or API keys you already use.

Similar tools let the Claude Code CLI or Codex CLI use other models, but they either cannot reach the Claude desktop app or reach it only through its gateway (3P) mode, and the moment you switch models you lose the Claude-subscription side (claude.ai chat, Remote Control, cloud sessions). ClaudeRipple covers the desktop app, the terminal and Codex from one menu-bar app, runs both subscriptions side by side, and keeps the Claude Code harness intact: your skills, hooks, MCP servers, CLAUDE.md, subagents and claude.ai connectors keep working while another model does the thinking.

¹ opencodex's README showed Claude Desktop in a demo but published no setup steps for it, in the repository or on its documentation site (checked 2026-09-16). At the time the only public way into the desktop app was the official gateway setting — the last column.

² CC Switch reaches Claude Desktop by writing the app's third-party profile (Claude-3p/claude_desktop_config.json, "inferenceProvider": "gateway"), which is the last column. Through its local gateway the picker shows only claude-sonnet-*, claude-opus-* and claude-haiku-* role names (its manual). CC Switch manages more tools than this table covers (Gemini CLI, OpenCode and others) and syncs MCP servers and skills across them.

³ Each ChatGPT account is bound to its own provider card, and official ChatGPT cards are kept out of the failover queue (v3.20.0 release notes).

The desktop app's own "third-party inference" setting flips the whole app into another mode: you lose claude.ai chat, Remote Control and cloud sessions. Tools that replace ANTHROPIC_BASE_URL never reach the desktop app, and the ones that do reach it write that same setting. ClaudeRipple instead is a tiny HTTPS proxy that only the Claude Code process trusts: requests for the models you map go to your provider, everything else goes to Anthropic byte for byte.

• Any model in Claude Desktop and Claude Code. GPT-6 Sol / Luna / Astra and GPT-5.6 Terra / Sol / Luna through your ChatGPT Plus/Pro subscription (new models appear the day OpenAI serves them), Google Gemini, or DeepSeek, Kimi, GLM, MiniMax, Qwen, Grok, Mistral, Groq, Together, Fireworks, OpenRouter (400+ models) and local Ollama / LM Studio. Under their real names in the picker, or mapped onto a Claude name.
• Claude in the Codex app and Codex CLI. A local OpenAI-compatible endpoint that Codex treats as a provider; Claude models appear in Codex's own model list. Uses your current Claude Code login or an Anthropic API key.
• Multiple Claude accounts without breaking the prompt cache. Opt a native Claude provider into account rotation, add subscriptions from the dashboard, and each conversation stays on the account that answered. Before the response starts, a quota or authentication refusal moves that turn to the next account. This native Claude Desktop/Code path is separate from the translated Codex ingress, which picks one available login and does not rotate accounts.
• Several ChatGPT accounts. Each clauderipple login (or + Add ChatGPT account in the dashboard) adds one. When an account hits its limit, the next one takes over the same request and the spent one rests until its window resets; a conversation stays on the account that answered, to keep its cache. This covers the Claude Desktop Code tab and subagents and GPT in the Codex app and CLI — Codex keeps its own sign-in and moves to the next account without signing out.
• The full Claude Code harness, untouched. Skills, hooks, MCP, CLAUDE.md, subagents, plan mode, Remote Control on your phone: nothing is turned off.
• Correct by construction. Prompt caching preserved (Anthropic cache breakpoints and stable OpenAI prefixes), Claude Code's server-side threads handled, tool calls and images round-tripped, reasoning effort clamped to what each model accepts, Anthropic-only request fields stripped for compatible vendors.
• A real request log. Who asked, which model answered, tokens in / cached / out, latency, status, per request, with a one-hour summary.
• Subagent name tags. The background-task panel shows Terra·high · Review instead of a generic "Agent".
• The ripple mod, above the prompt. A band shows the reasoning effort, context, cache timer, and which model answered your last request with its cache hit. A small judge model (Haiku 5.5 or Jev) reads each prompt with the recent turns and sets the effort, low through max. Delegated work goes to ripple:auto, and the judge sends each task to a worker by difficulty: light work to DeepSeek or GLM, hard work to GPT or Opus. Cards warn when the cache has gone cold, the context gets heavy or a plan window passes 80%, and offer compact or a handoff to a fresh session in the same folder. Log (or /ripple-log) opens the request log in a side pane. One switch on the Clients screen, or clauderipple mod on.
• Images from your ChatGPT subscription. clauderipple image "<prompt>", with reference images and transparent backgrounds. Subagents are told how, so a GPT or DeepSeek worker can make an image and look at it without an image tool.
• Close the laptop, Claude keeps working. On macOS and Windows the laptop stays awake with the lid shut while Claude works (and 10 minutes after), so a task you started from your phone finishes. If the Wi-Fi loses its internet, it moves to your phone's hotspot (iPhone or Android) and back to a known Wi-Fi later. Off until you turn it on in the dashboard's Laptop view; see Lid closed and the phone hotspot.
• Lean workers, if you want them. One switch gives generated subagents only the tools a worker uses instead of every MCP server and plugin of the session: one worker's first request fell from 33k to 9k tokens.
• Made for people who don't want a terminal. Provider presets with one-click connection tests and model discovery, drop-down model mapping, auto-save, a UI in English, Korean, Chinese, Japanese, Spanish and Portuguese. A menu-bar app that carries its own runtime and sets itself up on first launch. Signed and notarized.

Start a job in the Claude desktop app, close the laptop, and keep going from the couch. Ripple Remote is the phone app for ClaudeRipple: the same sessions, whichever model runs them, on your phone, and it does not drop you after 30 minutes.

• Hours, not minutes. The connection holds as long as your computer is on. Watch a two-hour refactor through to the end and approve it when it asks.
• Any model you route. GPT, DeepSeek, Gemini or Claude: every session looks and works the same. Switch the next turn's model and reasoning effort from your phone.
• Your Claude app sessions, as they are. Sessions you started in the Code tab on your Mac continue on the phone — the live timeline, tool calls, diffs, subagents and their models, plan approvals and question cards, photos and files.
• Your Mac's status bar in your pocket. The model that answered, the cache and your usage limits, as the ClaudeRipple band shows them on the Mac.
• Pairs with the laptop features. With lid-closed awake and the phone hotspot, the laptop keeps working in a bag and stays online when the Wi-Fi does not.
• Private by design. Messages, attachments and notifications are end-to-end encrypted; the relay and the push service cannot read them. No account, no ads, no tracking, free.

Install ClaudeRipple on your computer, open the dashboard's Phone view and scan the QR code once. iPhone: App Store. Android: in a closed test on Google Play — ask on Discord. Claude app sessions are supported on macOS; the app has a demo you can try without a computer.

macOS

curl -fsSL https://raw.githubusercontent.com/PBJ-2/clauderipple/main/scripts/install.sh | sh

Windows (PowerShell)

irm https://raw.githubusercontent.com/PBJ-2/clauderipple/main/scripts/install.ps1 | iex

Linux (a desktop session with systemd)

curl -fsSL https://raw.githubusercontent.com/PBJ-2/clauderipple/main/scripts/install.sh | sh

The router runs as a per-user systemd service: systemctl --user status clauderipple, and journalctl --user -u clauderipple for what it printed. Picker mode trusts the local CA in the NSS database Chromium reads (~/.pki/nssdb, your user only, no sudo), which takes certutil: run sudo apt install libnss3-tools (Fedora: sudo dnf install nss-tools) before clauderipple picker on.

That is the whole prerequisite list. The script uses a Node 24+ you already have, and downloads the official build into ~/.clauderipple/runtime when you have none, verifying it against the checksum nodejs.org publishes. Then it sets ClaudeRipple up: a local certificate, two lines in ~/.claude/settings.json, and a background router that starts with your computer. No administrator rights, no password.

With Node already installed you can skip the script:

npm install -g clauderipple
clauderipple install

Then open the dashboard and add a provider:

clauderipple ui

Providers → add ChatGPT or paste an API key → Model mapping.

The menu-bar / tray app is optional. It shows the router's state and opens the dashboard in a click:

clauderipple tray

It runs on Electron, which is about 270MB and is therefore not installed by default. clauderipple tray --install fetches it once; everything else works without it.

Updating. The tray offers a new version when there is one (Check for Updates…), or:

clauderipple update

It updates the way you installed — the install script again, or npm — and restarts the router on the new version. The tray's Electron is kept across updates.

Optional, for real names in the Desktop picker: Clients → Claude Desktop → Model picker → on, then quit and reopen Claude Desktop. To trust the local certificate for your user only, macOS asks for your login password and Windows shows a confirmation dialog with the fingerprint; Linux adds it to your NSS database without asking. ClaudeRipple never sees a password, and no platform needs administrator rights.

Closing Claude Desktop's window is not enough. It keeps running, and the next launch reuses it without reading the new setting. Quit it properly (macOS: ⌘Q; Windows: the tray icon, or Task Manager; Linux: Ctrl+Q or the tray icon, until pgrep -f claude-desktop prints nothing) or the picker will silently not change.

git clone https://github.com/PBJ-2/clauderipple && cd clauderipple && npm install
node packages/cli/src/index.ts install   # certs, settings.json env, supervisor (launchd / Task Scheduler), end-to-end probe
node packages/cli/src/index.ts ui        # open the local GUI in your browser

Works out of the box after install. Map models on Model mapping (auto-saves), or turn on picker mode to see provider models by name in the app's own picker. Effort chosen in the app is passed through; unsupported levels are clamped.

Same router, same mapping. /model gpt-5.6-terra lists the models you added. Subagents follow the routing; a [[gpt: sol@xhigh]] marker in a subagent prompt overrides the model for that call, and the task panel shows the real model name.

You do not have to build an agent harness for this. Point the subagent slot at a routed model and every subagent runs on it — no agent file, nothing else:

// Settings → CLI models, or "cli": { "models": { … } } in the config
{ "subagent": "gpt-5.6-terra" }   // → CLAUDE_CODE_SUBAGENT_MODEL

Every model you tick is also a named subagent. The router writes ~/.claude/agents/<name>.md for each model exactly one provider carries (gpt-5.6-terra → gpt-5-6-terra, deepseek-v4.1-flash → deepseek-v4-1-flash), so "have DeepSeek do it" needs nothing beyond ticking the model: the Agent tool lists it, subagent_type: "deepseek-v4-1-flash" runs it, and the task panel names it. Untick the model and its file goes away. The router only ever touches the files it wrote (recorded in generated-agents.json); an agent file you wrote yourself under the same name wins and is left alone, so a custom brief is still just Claude Code's own agent file with a model: the router knows:

---
name: reviewer
description: Independent review on GPT-5.6 Sol.
model: gpt-5.6-sol@medium
---
You are the reviewer for this session. Verify the change yourself and report
the conclusion only.

Per-call effort goes on the first line of the subagent prompt: [[ripple: gpt-5-6-terra@high]] — any agent name or model id works there. A marker anywhere else is prose and is ignored (a compaction summary that quotes one must not re-route the session). A model the router cannot place is refused by name (400 ClaudeRipple: no provider declares "…") instead of being sent on to fail somewhere less legible; native Claude models always pass through. Set "cli": { "agentFiles": false } to turn generation off.

By default a generated worker inherits every tool of the session, so each MCP server and plugin rides along on every worker request. "cli": { "limitWorkerTools": true } (or Worker tools in the dashboard) gives it only the read, edit, run, search and skill tools, which cut one worker's first request from about 33k to 9k tokens. Such a worker has no browser or simulator tools. A worker that needs them belongs in an agent file of your own, which is never touched.

With a ChatGPT provider configured, images come from the same subscription:

clauderipple image "a red paper boat, flat illustration" -o boat.png --aspect square

--transparent, --format png|jpeg|webp and --ref FILE (a reference image, repeatable) are optional. One image takes about 30 seconds. The subscription backend picks the resolution and quality itself, so the shape is asked for in words (--aspect). The command is a client of POST /api/image on the admin port, which answers with the image bytes. Generated workers are told the curl form of it, so a worker with only Bash can make an image and read it back.

Claude Code runs WebSearch as a separate small-model request. To keep a ChatGPT-routed setup from spending Anthropic quota for that request, select the same ChatGPT provider as the search backend in config.json:

{ "webSearch": { "provider": "chatgpt", "model": "gpt-5.6-terra" } }

This is opt-in. With no webSearch setting, Claude Code's existing search path is unchanged.

Claude Desktop's Code tab can run a session on a remote Linux host over SSH. Claude Code then runs on that host, where ClaudeRipple is not, so routed models are unknown there (Agent type 'gpt-6-1-sol' not found). Connect the host once from the machine ClaudeRipple runs on:

clauderipple remote add my-server            # an ssh alias or user@host; --port N (default 18791)
clauderipple remote hosts                    # tunnel state per host (also in `remote status`)
clauderipple remote remove my-server         # undo it

remote add needs key-based ssh my-server to work without a password prompt. It copies the router's CA certificate (the public certificate only) to ~/.clauderipple-remote/ on the host, points the host's ~/.claude/settings.json at the tunnel (HTTPS_PROXY=http://127.0.0.1:18791, NODE_EXTRA_CA_CERTS; the original is kept as settings.json.clauderipple-backup; a different proxy already there is refused without --force), and copies the generated agent files. The router then keeps a reverse SSH tunnel open from the host's 127.0.0.1:18791 to itself, restarts it when it drops, and copies the agent files again whenever they change. The host needs only sh — no Node and no ClaudeRipple.

While the tunnel is down, Claude Code on the host gets connection refused rather than going to Anthropic around the router. The remote session only exists while this machine's Claude Desktop is connected to it anyway. The tunnel listens on the host's loopback only, but on a shared host every user there can reach that port, and through it your router. The status mod and image generation are not available on remote hosts yet.

For sessions you drive from your phone (Remote Control, Ripple Remote): the laptop keeps working in a bag. Open the dashboard's Laptop view.

Stay awake with the lid closed. Three modes: Off; While working — awake while a model call runs and for 10 minutes after the last one; Always — until you turn it off or a time you set. It lets go by itself when the battery drops below 20% or (on macOS) the Mac reports a thermal warning, and puts a closed laptop to sleep when it does. Don't carry it in a closed bag for long while it is awake.

• macOS: press Set up once. macOS's own password dialog installs a sudoers rule for exactly the commands used (pmset disablesleep, pmset sleepnow, joining the hotspot you name); ClaudeRipple never sees the password. clauderipple awake setup --remove takes it out. If Insomnia is installed, setup removes it (it does the same job).
• Windows: no administrator rights. A child of the router holds SetThreadExecutionState, and the lid action is set to "do nothing" only while it is held; your own power settings come back afterwards, also if the router dies.

Wi-Fi without internet → phone hotspot. It probes the internet through the Wi-Fi interface (a router with a dead uplink or a captive portal still looks "connected" to the OS) and joins the hotspot you name after 3 of 5 rounds fail. On the hotspot it tries the way back 5, 10, 20, 30, 40, 50 and 60 minutes after landing, then every 10 minutes: it leaves the hotspot so the OS joins the best known Wi-Fi in range, and stays there only if that Wi-Fi has internet. A model call that starts during the switch waits a moment instead of failing.

• Join the phone's hotspot from the laptop once by hand, so the OS keeps its password.
• iPhone: keep Settings › Personal Hotspot › "Allow Others to Join" on. On a Mac, the same Apple Account and Bluetooth on help it find the phone.
• Android: an idle hotspot turns itself off after a while; turn it on again if it is not found.
• Linux: not supported yet.

clauderipple codex on     # adds a "clauderipple" provider to ~/.codex/config.toml (backup first)
codex --profile clauderipple -m claude-sonnet-5

codex on also points Codex's built-in OpenAI provider at ClaudeRipple (openai_base_url). Codex keeps its ChatGPT sign-in; its GPT requests go on to ChatGPT unchanged, except that the account is chosen from the ones added to ClaudeRipple — so with several added, Codex too moves from a spent account to the next. With none added, or all of them at their limit, Codex's own login is used. An openai_base_url you set yourself is left alone. clauderipple codex off undoes it.

Claude models show up in Codex's model list under their names (ClaudeRipple writes a model catalog next to Codex's own). Claude is reached through your Claude Code login (detected from the running Desktop session, the terminal login, or a sign-in of ClaudeRipple's own: Providers → Claude (Anthropic) → + Add Claude account opens the browser, no terminal needed; the same as clauderipple claude-login) or an Anthropic API key. Reusing a subscription login is subject to Anthropic's terms. Any Anthropic-compatible provider you configured is available the same way.

Requests to compatible providers are cleaned of Anthropic-only fields (server-side threads, deferred tools, context management, thinking bindings) and effort is clamped per model, so vendors do not 400 on Claude Code's request shape.

Claude Desktop / claude CLI ──HTTPS_PROXY──▶ ClaudeRipple ──▶ api.anthropic.com   (unchanged)
                                               │
                        mapped model ──────────┼──▶ chatgpt.com/backend-api (Responses ⇄ Messages)
                                               ├──▶ Anthropic-compatible vendors (+ compat layer)
                                               ├──▶ OpenAI-compatible vendors (Messages ⇄ Chat/Responses)
                                               └──▶ Google Gemini (Messages ⇄ generateContent)
Codex app / CLI ──/v1/responses──▶ ClaudeRipple ingress ──▶ Claude (your login or API key) / vendors

• The Claude Code CLI reads HTTPS_PROXY and NODE_EXTRA_CA_CERTS from ~/.claude/settings.json (Anthropic's documented corporate-proxy path). Only that process trusts ClaudeRipple's local CA; the OS keychain is untouched unless you turn on picker mode.
• Picker mode routes the app's own claude.ai traffic through the proxy and adds your models to the picker list the app fetches at start. Off again with one click.
• The router drains in-flight calls on restart, self-exits on repeated upstream failures so launchd restarts it, rotates logs, and never drops a request silently.

Details with sources: docs/ARCHITECTURE.md.

Everything runs on 127.0.0.1. API keys live in ~/.clauderipple/config.json and added Claude OAuth grants in ~/.clauderipple/claude-accounts.json (both 0600). Admin APIs and logs expose neither tokens nor upstream account ids. The only network destinations are the providers you configure. There is no telemetry.

In daily use by the author, but young. Expect rough edges:

• Windows support is new (2026-09-14). Install, setup, the router, signing in, the picker and a live GPT call are verified on x64 hardware; crash recovery on arm64. The window still has the stock Windows title bar. Windows builds are unsigned; see the note under Install.
• Linux support is new (2026-09-24), contributed with measurements on Ubuntu 24.04 x64 (install, a Code-tab request, restart mid-stream, picker mode, uninstall). It needs a systemd user session; other distributions and arm64 are untested.
• ChatGPT, OpenRouter and Claude-in-Codex are verified with live accounts; the other presets follow the vendors' official documentation.
• A model added to the picker is usable from the next session.
• Claude Code and Codex change their wire protocols often; a client update can break a translation until ClaudeRipple catches up. Bugs and logs welcome.

Questions, bug reports and the Ripple Remote Android beta (a closed test on Google Play) are on Discord: https://discord.gg/R8ZgGSNHjw. To join the Android beta, leave the Gmail address you use on Google Play there.

ClaudeRipple is an independent open-source project. It is not affiliated with, endorsed by, or sponsored by Anthropic or OpenAI. Claude and Claude Code are trademarks of Anthropic, PBC. ChatGPT and Codex are trademarks of OpenAI.

Copyright (c) 2026 pbj. GPL-3.0 — see LICENSE. Use it freely; if you distribute a modified version, ship its source under the same license.