Lend your AI wIthout giving away your keys.

The open-source project tokenbud released a desktop app that lets a host share access to a local or API-backed AI model through a 9-digit ID and password, without handing over the underlying API key. Hosts set a cap on concurrent users and an hourly prompt limit, and can revoke access by changing the password; prompts route through a Cloudflare Worker relay that stores only a password hash and deletes finished prompts after 10 minutes. The app supports local models in Ollama or LM Studio and API keys for Anthropic, OpenAI, Google Gemini, Mistral, Groq, DeepSeek, xAI, OpenRouter, or any OpenAI-compatible API, and its builds are not yet code-signed.

Lend your AI. Keep your keys. You pay for a great AI, or you run one on your own computer. Now a friend, a teammate or a whole class wants a turn. tokenbud lets them, with an ID and a password. No key changes hands, no account is shared, and you can end it with one click. Download https://github.com/EagleMind/tokenbud/releases/latest · How it works how-it-works · Run your own relay run-your-own-relay · Build from source build-from-source Most of the time, the AI you pay for sits idle. Meanwhile someone you know is stuck without one, or paying twice for something you already have. Until now, helping meant one of two bad choices: hand over your API key, the secret that unlocks your account and your bill, or set up a server and hope you got the security right. tokenbud is the third way. Open it, press Share my agent , and you get a 9-digit ID and a password. Read them out, text them, write them on a sticky note. The other person types them into tokenbud, or into a web page with nothing to install, and starts chatting with your AI as if it were their own. You stay in charge the whole time: - You set the limits. Choose how many people can chat at once and how many questions you'll lend each hour. When the hour's budget is spent, sharing pauses on its own and they can see when it comes back. - Your key never leaves your computer. They get answers, never the secret behind them. Their questions are answered on your computer, by your model or your account. - One click ends it. Kick one person, or change the password and everyone is out. - You don't set anything up. tokenbud connects through a ready-made go-between that the project runs for you around the clock, so it works behind any home router. Prefer your own? You can run your own run-your-own-relay in a few minutes. Think of it as a library card for your AI: people borrow time, you keep the book. - Simple to share. A 9-digit ID and a password. A new password disconnects everyone. - Seats you control. Choose how many people can prompt at the same time; everyone else waits for a free seat. - A reserve that protects you. Lend at most N prompts an hour. Sharing pauses on its own, and people see when it resumes. - Your model, your choice. A local model in Ollama or LM Studio, or your own key for Anthropic, OpenAI, Google Gemini, Mistral, Groq, DeepSeek, xAI, OpenRouter, or any OpenAI-compatible API. - Nothing to install for guests. The default relay also serves a web client, and the desktop app does the same job. - Local-first memory. Chats and the facts you want the agent to know stay on your device, and are sent along only while you're connected. - Agent to agent. tokenbud mcp lets your own agent Claude Code or any MCP client prompt a shared one. - Light and dark, with a calm, flat look. | <sub Chat with someone's agent. It can suggest facts to remember, and nothing is saved until you say so.</sub | <sub Pick what answers: a local model, or a provider with your own key.</sub | guest app or browser ──HTTPS──▶ relay ◀──WebSocket── tokenbud host ──▶ your model ◀─answer── Cloudflare Worker ──answer───────▶ local or API 1. The host app connects out to a small relay on Cloudflare and gets an ID. No open ports, no port forwarding, works behind any router. 2. A guest opens a session with the ID and password. The relay checks the password it only stores a hash and gives them a seat if one is free. 3. Prompts travel through the relay to the host, which answers with its model and sends the answer back. Finished prompts are deleted from the relay after 10 minutes. Each host has its own Durable Object on the relay. It holds the password hash, the seat count and the live sessions, and nothing else. Get the latest installer for Windows, macOS or Linux from Releases https://github.com/EagleMind/tokenbud/releases/latest . The builds are not code-signed yet, so Windows SmartScreen and macOS Gatekeeper ask for confirmation the first time. Open tokenbud and click Share my agent . Sharing starts right away and shows your ID and password. - People at the same time: set the number of seats 1–50, default 3 . You see who's connected and who's prompting, and can disconnect anyone. - Copy link gives guests the web client with your ID filled in; they only type the password. - ↻ makes a new password and disconnects everyone. The switch at the top stops sharing. Click Connect to an agent , or open the relay's address in a browser. Enter the ID and password and start chatting. Start as many chats as you like; follow-ups keep their context. Closing the window or tab gives your seat back. The ⚙ button opens Settings: - Model. Where answers come from see below . - Reserve. Lend at most N prompts an hour default 30, rolling . It also shows what you've lent. - Relay. The default relay a Cloudflare Worker run by the project , or your own. - Appearance. Light, dark, or match the system. The window opens at one size for both modes, remembers the size and position you leave it at, and the layout adapts to the width. | Choice | What it uses | Notes | |---|---|---| | This computer | Ollama https://ollama.com , LM Studio https://lmstudio.ai , or any OpenAI-compatible server | Found automatically on their usual ports. Free and private. | | A provider | Anthropic, OpenAI, Google Gemini, Mistral, Groq, DeepSeek, xAI, OpenRouter, or Custom | Your own API key, one per provider, kept in the OS keychain. The model list comes from the provider. | | Claude Code | Your own Claude Code login | Development builds only, for testing on your own machine. See Using it within the rules using-it-within-the-rules . | Answers from reasoning models arrive without their <think sections. Your chats and a short list of facts about you "Writes Rust", "Prefers short answers" are kept only on your device : in the app's config folder, or in the browser's IndexedDB. - Facts travel with each prompt to the agent you're connected to, so it answers knowing you. - The agent can suggest a fact , which appears as a Remember / Not now chip. - Open Memory in the sidebar to add, edit or delete facts, or to switch memory off. The same executable is also an MCP server, so Claude Code or any MCP client can prompt a shared agent: claude mcp add tokenbud -e TOKENBUD ID=418203977 -e TOKENBUD PASSWORD=k7m2pqxa -- tokenbud mcp It adds one tool, ask prompt . Each call takes a seat only while it runs. tokenbud uses the project's relay by default, a Cloudflare Worker that is kept running for everyone. You can also run your own on Cloudflare's free plan in a few minutes, and then only the people you choose share it. You need a free Cloudflare account https://dash.cloudflare.com/sign-up and Node.js https://nodejs.org 20 or later. git clone <this repository tokenbud cd tokenbud/desktop && npm install the web client is built from here cd ../relay && npm install npx wrangler login opens the browser once npm run deploy builds the web client, then deploys the Worker The deploy prints your relay's address, such as https://tokenbud-relay.<you .workers.dev . Then: 1. In tokenbud, open Settings → Relay , choose My own relay , paste the address, and click Use this relay . The app checks that it's a tokenbud relay before switching. 2. Everyone you share with, and everyone you connect to, has to use the same relay. Guests can also simply open the relay's address in a browser. 3. You get a separate ID on each relay. Your ID on the default relay is kept. Good to know: - Free plan limits. The Workers free plan allows about 100,000 requests a day. A connected guest checks in about every 15 seconds, so that's roughly 15 to 20 people connected all day. The $5 paid plan removes the worry. - Custom domain. Add one in the Cloudflare dashboard if you'd like a nicer address. - Environment override. TOKENBUD RELAY=https://… overrides the setting, which is handy for scripts and testing. - Updates. Run npm run deploy again after updating tokenbud. To make your relay the default in builds you hand out, change DEFAULT RELAY URL in crates/tokenbud-core/src/lib.rs . You need Rust 1.90 or later and Node.js 20 or later. On Linux, also the Tauri prerequisites https://v2.tauri.app/start/prerequisites/ . cd desktop && npm install && npm run build && cd .. cargo build --release The app is target/release/tokenbud . To make installers, run npm run bundle in desktop/ . For a release on GitHub, push a version tag: the release workflow https://github.com/EagleMind/tokenbud/blob/main/.github/workflows/release.yml builds Windows, macOS and Linux installers and attaches them to a draft release. git tag v0.1.0 && git push origin v0.1.0 tokenbud never sees or moves a credential. But model providers decide what their accounts may be used for, so here's how each choice fits: - Local models involve no provider at all. - Provider keys are meant for building on, including for other people. You're responsible for how your key is used, and providers forbid reselling access, so share, don't sell. - Claude subscriptions Free, Pro, Max are for individual use. Anthropic's terms don't allow answering other people's prompts with them. That's why the Claude Code engine exists only in development builds. To share Claude, use an Anthropic API key. - The relay stores only hashes: salted PBKDF2 for passwords, SHA-256 for host secrets and session tokens. Ten wrong passwords in a minute from one address lock out that address only, so nobody can block a host's other guests. - An unknown ID and a wrong password get the same answer, so IDs can't be probed. - Web client tabs keep only their session token, never the password. A refresh resumes the same seat; closing the tab frees it within 45 seconds. - The desktop app and the web client have a Content Security Policy, so answers can't load outside images or scripts. - The relay sees prompts and answers in transit and deletes finished ones after 10 minutes. Run your own relay if that matters to you. - Keys and secrets live in the OS keychain Windows Credential Manager, macOS Keychain, Linux keyring . - With a local model, prompts never leave the two computers and the relay. | Path | What it is | |---|---| | apps/desktop | tokenbud , the desktop app Tauri 2 : host and client modes, and tokenbud mcp | | apps/tokenbud-host | Sharing: the relay connection, seats, the reserve, and the model engines | | apps/tokenbud-client | Connecting by ID and password, and the MCP server | | crates/tokenbud-core | Wire types, relay client, settings and keychain storage | | relay/ | The Cloudflare Worker: web client at / , API at /v1 | | desktop/ | All the UI: React, Tailwind and shadcn/ui. src/app is the desktop window, src/web the browser client | Development cd relay && npm run dev relay + web client on :8787 cd desktop && npm run app the desktop app with hot reload cd desktop && npm run dev:web web client with hot reload on :1422 proxies /v1 to :8787 cd desktop && npm run preview:ui the app's UI in a browser with sample data Run the app against the local relay with TOKENBUD RELAY=http://127.0.0.1:8787 . - TOKENBUD HOME=/some/dir keeps an instance's settings in that folder, so several can run on one machine. - cargo run -p tokenbud-host --example share shares from a terminal and prints the ID and password. It's useful for testing without the window. In the UI preview, ?steps=host,settings:model clicks through to a screen, ?theme=dark forces dark mode, and ?connect=1 shows the connect screen. The README screenshots are taken that way. Relay API | Method and path | Auth | Purpose | |---|---|---| | POST /v1/hosts | none | New host: returns host id 9 digits and host secret . | | GET /v1/hosts/:id/connect WebSocket | host secret | The host's channel. | | POST /v1/hosts/:id/sessions {password, name} | none | Take a seat. Returns session token . 401 wrong ID or password, 409 no free seat, 429 too many tries, 503 not sharing. | | DELETE /v1/hosts/:id/sessions | session | Leave. | | GET /v1/hosts/:id/ping | session | Keep the seat every 15 s; 45 s without one ends the session . Returns online and paused . | | POST /v1/hosts/:id/ask {prompt, history, context?} | session | Send a prompt. context carries the guest's memory at most 10 KB . 409 if one is already running, 503 while paused. | | GET /v1/hosts/:id/jobs/:job?wait=25 | session | The answer, with long-polling. | | GET /health | none | {"service": "tokenbud-relay"} | Errors are JSON: {"error": "A sentence for people.", "code": "wrong password"} . Decide on code , never on the sentence. The codes are wrong password , rate limited , seats full , busy , not sharing , paused , session ended , too large , bad request and not found . Request bodies are limited to 1.5 MB. On the WebSocket: - The host sends config {password, seats, name} , pause {paused} , kick {session id} and result {job id, status, text?, error?} . - The relay sends sessions , job and cancel . tokenbud is a hobby project, maintained when time allows. Issues and pull requests are welcome; small and focused ones are easiest to review. CONTRIBUTING.md https://github.com/EagleMind/tokenbud/blob/main/CONTRIBUTING.md has the setup, the checks CI runs, and the code style. Changes are listed in the changelog https://github.com/EagleMind/tokenbud/blob/main/CHANGELOG.md . MIT https://github.com/EagleMind/tokenbud/blob/main/LICENSE . Model providers' own terms still apply to the keys and accounts you use with it.