01 — Built for founders who pay for tokens they shouldn't
Built for founders who'd rather ship than paste keys into five dashboards.
One click signs you up across the automatic half of 20 free LLM providers, collects every key onto your Mac, and puts around 400 verified chat models behind one local key — at least 162 million free tokens a month, renewed monthly. You chat, ask about your files, and press Apply on every edit.
02 — How it works
Walk a sample vault through harvest, a chat on those keys, and an agent with a playbook. Preview Get Tokens or an Apply. Everything stays in this browser. The same Chat is what you open after you install.
Saved in this browser. Sample vault. Not your keys. This is the Mac app, in this page.
03 — What it does
FreeTokens is a native Mac app that does the boring half of using free LLM tiers: the twenty signups, the twenty dashboards, the twenty keys. You press one button and get a working key pool, on your own machine.
http://localhost:8787/api/v1. The app writes each tool's config for you and then checks the endpoint answers on it.04 — How much you get
Twenty providers hand out a real free tier each. On their own they are twenty accounts you would never keep track of; collected in one pass they are a single pool one key can spend — and when one runs dry the gateway routes to the next without telling you about it. The numbers below are the app's own: it derives them from its directory rather than carrying them as copy, so they cannot quietly go stale.
~162M+ tokens a month
The sum of what 7 of the 20 providers publish — each one's own daily token cap, or its request cap at 2,000 tokens a turn. The plus is the other 13: signup credit, tiers behind a phone number, endpoints metered by IP, and one that publishes nothing.
~400 models you can chat with
The shipped catalogue verifies 396 chat models today — embedders, rerankers, safety classifiers, speech and image models are filtered out, because none of them can hold a conversation. Update Models then re-asks every provider you hold a key for what it serves now, so the figure on your machine is yours rather than ours.
10 of 20 walked end to end
2 more stop once for the step a bot cannot do. 6 have no agent path and are a paste-the-key job. 2 need no key at all.
$0 to run the harvest
A Google AI Studio key takes about a minute and asks for no card, and its free tier covers a full harvest several times over.
~$2 if you prefer Claude
About $0.23 a provider on Haiku with the prompt cached — roughly $2 for the whole automatic set, billed to your own Anthropic key.
4,000 characters read
Waiting on a code it searches the last two hours, opens the one message from that provider, and reads the first 4,000 characters. Nothing else is opened, nothing is copied to disk.
Limits are set by each provider and change without notice, so nothing here is carried as a promise. Update Models proves every candidate with a real request on your own key before it is added, and retires the names a provider has stopped serving; Usage shows what is left of today's allowance while you spend it. Mistral stopped publishing a free limit altogether, so it sits in the plus rather than in the total — and Gemini's share is the figure this app ships rather than Google's current word, which it no longer gives.
05 — Harvest, vault, chat
Whether the vault is empty or you already have keys on disk, you start here. Get Tokens collects what it can. Manual providers wait for you. Then you talk in Chat, with folder and files on the thread.
Press Get Tokens. The app opens each of the ten automatic providers in its own browser and collects a free key; one whose key still works is skipped. Two more get as far as the single step a bot cannot do — a CAPTCHA, a card — and stop to ask you, in the app. Six have no agent path at all, because a phone code or a consent screen blocks it: for those the app opens the key page and you paste the key back in. Two need no key in the first place.
Collected keys live in the macOS Keychain. Older ~/.freetokens/keys.json files import on first run. Copy a key when you need it. It does not leave this Mac unless you copy it.
One Chat. Folder and files are context, not destinations. Auto-route across harvested keys, or pin Direct to a vault provider. Named agents open in the same Chat with a playbook attached. Diffs sit in the thread until you press Apply.
The app's own tour lists these, and says beside each one which key it needs. 5 run on whatever the harvest brings back; the other 4 each want one particular provider's key, so you can see what you are unlocking before you start.
Chat that picks a provider for you and falls back when one is down.
Ask about things that happened after the model was trained; answers cite sources.
Drop in PDFs, notes or code and ask questions across all of them at once.
A saved assistant with its own instructions and its own files, shareable by link.
Say “draw me…” in chat and it makes the picture — no separate screen.
Photograph a page or a receipt and ask what it says.
Talk instead of typing; you get text back to check before it sends.
Pick one provider and one model by hand and see exactly what it returns.
Point Claude Code, Codex, Zed or any OpenAI client at one local endpoint.
06 — Agents
An agent is a set of standing instructions, a folder, its own documents and its own SKILL.md playbooks. Opening one lands in the same Chat with all of that already on the thread — and edits still wait for Apply.
Answer from the vault. Auto-route across harvested keys. No extra playbook — this is Chat with nothing pinned to it.
Read the Lovie deck. Be blunt about what a partner would skip. Do not rewrite the company.
Folder, files and playbooks ride with the agent. Create stays on the roster. Diffs still wait for Apply.
07 — Native Mac
FreeTokens is a Swift app. No bundled Chromium. Harvest runs in the app's own browser, the gateway runs on localhost, and the vault lives in your Keychain. The only thing on the wire is your own call to the provider — your machine, their server, nothing of ours in the middle.
Native macOS. Signed and notarized, so it opens without the Gatekeeper detour. Drag the dmg to Applications.
Every request goes from this Mac straight to the provider that issued the key. Lovie runs no proxy and hosts no vault — we never see a token, a prompt or an answer.
Waiting on a verification code, the agent searches the last two hours, opens the one message from the provider it is working on, and reads the first 4,000 characters. That text goes to the model driving the harvest — the one running on your own key — and nowhere else. Nothing is stored.
08 — Download
Signed and notarized. Open the dmg, drag the app to Applications, sign in with your Lovie account. It needs macOS 14 Sonoma or later, and the build below is whatever GitHub is serving as the latest release right now.
One universal dmg — both architectures in the same file, so there is no build to pick.
Apple Silicon and Intel · notarized
09 — Questions
No. This page is a sample with fake keys. Get Tokens runs in the Mac app, in its own browser, on your machine — and only for the 10 providers the agent can walk end to end. Two more stop to ask you for one step, six are a paste-the-key job, and two need no key at all.
On this machine, in the macOS Keychain. Not on Lovie servers. An older ~/.freetokens/keys.json from the Electron app is imported on first run, and your conversations sit next to it in ~/.freetokens/chats.json.
Sign in with your Lovie account to open the app. You still sign up at the providers for the tokens themselves, and you bring one free key of your own to drive the agent — a Google AI Studio key takes about a minute and asks for no card. Lovie does not issue any of them.
Both. The dmg carries one universal build — Apple Silicon and Intel in the same file — so there is no version to choose between; it needs macOS 14 Sonoma or later.
The dmg is signed and notarized, so it opens by double-click. Open it and drag the app to Applications. If macOS nags anyway, the download came from somewhere other than the button on this page.
Only when you press Apply, in the Mac app. This landing demo never writes to disk. Diffs stay in the thread until you say so.
Standing instructions plus a folder, documents and SKILL.md playbooks, all on the same Chat. Opening one lands in Chat with its folder, files and playbooks attached. Create stays on the roster. You press Apply on edits.
No — that is the other half of the same app. FreeTokens is the vault, the gateway and Chat; Accelerator AI answers your founder profile once and fills in accelerator applications from it. One download will carry both — Accelerator AI has not shipped yet, and there is a page for it with a demo of its own.
Sample vault above. Real keys in the app. Apply stays with you.