One email. No follow-up unless I ship something that fits.
Read firstFree tokens are not frontier models. This sends Claude Code through free tiers of other AI providers. You keep the Claude Code tool, but the model answering is whichever free one the router picks. Expect weaker answers and slower replies than a paid plan.
WHAT YOU ARE BUILDING
Claude Code, a local router, then free providers
Three parts, all on your own computer except the providers.
the path of one requestplain words
Claude Codeasks http://localhost:3001 (the FreeLLMAPI router)FreeLLMAPIpicks a free model from the providers you added keys forA free tierat Google, Groq, Cerebras and others, using your own free keys
FreeLLMAPI's README says it pools free tiers from 34 providers, and estimates up to 7.4 billion free tokens a month across them. That is a ceiling for the whole catalogue, not what you get. Your real allowance is whatever the providers you sign up to give you.
BEFORE YOU START
Four things to have ready
Claude Code installed. Run claude --version to check.
A free key from two or three providers. Sign up on each provider's own site and create an API key.
Ten minutes, and one of three install paths below.
Mac or Windows: the desktop app is simplest. Linux or a server: use Docker or Podman, or Node.
THE SETUP CHECKLIST
Seven steps, in order
Step 1Install the routerPick one path, shown in the next section. Wait until the dashboard opens at http://localhost:3001.
Step 2Create your accountThe page says "Create your account". On the same computer, just do it. From another device, enter the first run setup code printed in the server log.
Step 3Add provider keysOpen the Keys page and add two or three providers from the starter list below. Paste each key.
Step 4Copy the unified keyThe Keys page header shows one key starting with freellmapi. Your apps use this and never see your provider keys.
Step 5Preview the changeRun the dry run below. It shows what would be written and changes nothing.
Step 6Connect Claude CodeRun the same command without the dry run flag, or use the launch command to keep credentials out of config files.
Step 7Send a small testStart claude, ask for something tiny, then watch the request appear in the dashboard.
ShortcutThe dashboard also shows a Getting started card with three steps: add a provider key, send a test request, connect your app.
STEP 1: THREE WAYS TO INSTALL
Pick one
Desktop app. On the GitHub releases page, take the macOS dmg or the Windows exe. No account or password is needed, and the only credential is the unified key from the tray popover. The README names these two.
Docker one liner. Needs Docker or Podman with Compose. It writes a folder called freellmapi in your home directory and listens on localhost only.
The README says you can read the script first at freellmapi.co/install.sh. Do that if you do not pipe web scripts into bash. Re running it keeps your .env and encryption key.
Run it from source with Node. The package asks for Node 20.18 or newer but below 25. Open the dashboard on port 5173 in this mode.
terminalNode route
git clone https://github.com/tashfeenahmed/freellmapicd freellmapi && npm install# create .env with ENCRYPTION_KEY, see .env.examplenpm run dev
Catch I hitOn my Linux machine npm install skipped one optional part, better-sqlite3, because it could not build it on Node 20. The server then stopped with better-sqlite3 is not installed. Node 22 had a ready made build and worked. Docker avoids this, but I had no Docker to run it.
STEPS 5 AND 6: PREVIEW, THEN CONNECT
What the setup command writes
This is a real dry run from my machine on 25 Oct 2026, with my key replaced. Nothing was written.
Read it: Claude Code will talk to your router and ask for the model called auto, so the router chooses. The README says the setup command backs up your existing config and merges into it.
Apply it: run the same line without --dry-run. Or skip the file:
terminalno config written
npx freellmapi launch
Before any setup, npx freellmapi doctor claude told me "No configuration found. claude will use its vendor default endpoint, not this gateway." Run it again after setup to see the change.
IF IT SAYS NO PROVIDERS
The first error you may see
If you ask for a reply before adding a key, the router answers with this. It is the most common first error. I got it on my machine on 25 Oct 2026.
router replybefore adding keys
no_providers_configuredNo candidate model has a configured, usable provider key. Add provider API keys in the dashboard.
The fix is Step 3. Add a key on the Keys page and send the test again.
WHICH PROVIDERS TO ADD FIRST
Eight with no catch listed
The project docs mention no special catch for these. They need an ordinary free key. This shortlist is my choice from the docs, not the project's ranking. Check each provider's own current terms before you sign up.
Google Gemini
Key from Google. Native Gemini adapter.
Groq
Plain key.
Cerebras
Plain key.
Mistral
Plain key.
OpenRouter
Models ending in :free cost nothing.
NVIDIA NIM
Plain key.
Cloudflare Workers AI
Enter the key as account_id:token.
Cohere
Plain key.
Replies are not equalProvider quotas reset daily at UTC midnight per the README, and quality can drop as the caps are hit. Add more than one provider so the router has somewhere to go.
PROVIDERS WITH A CATCH
Skip or read carefully
All of these are in the project's own provider notes. They change often, so read the current list.
OpenCode Zen
Its free models are locked to the OpenCode client since September 2026 and fail with a 403.
ModelScope
Needs an Alibaba Cloud China account with real name verification.
Kilo, no key
Limited to 200 requests an hour per IP. Prompts and outputs are logged for training.
OVHcloud, no key
About 2 requests a minute per IP.
Hugging Face
A US$0.10 monthly credit, so it runs out fast.
SambaNova, Reka, Sail
No free tier left, or a payment method is needed. SambaNova ended its free tier in June 2026.
Nara
Free plan asks for Telegram verification.
PrivacyDo not send client data, passwords or private code through free tiers unless you have read that provider's terms on logging and training.
THE HONEST LIMITS
What free costs you
No frontier models. Expect a good open model, not the best one.
Speed varies by provider and by the hour. There is no service level promise.
Quality can dip late in the day as daily caps are hit. They reset at UTC midnight.
It is a single user router for personal experiments. Do not run client deadlines on it.
Free plans change without notice. A provider that works today may not next month.
Free installs get the model catalogue about 30 days late. The README sells a live catalogue at US$19 a year. You do not need it to start.
KEEP IT SAFE
Four rules
The router listens on localhost only by default. Leave it that way. Setting HOST_BIND=0.0.0.0 opens it to your whole network, which the docs say to do only on a trusted LAN.
The unified key is like a password for your free quota. Do not paste it in public chats or commit it.
Your provider keys are stored encrypted. Keep a copy of the .env file, which holds the encryption key.
To undo setup, restore the backup it made of ~/.claude/settings.json or delete the env block it added. The launch command leaves nothing behind.
THE ONE LINE CHECK
Check Claude Code is really using it
terminalany time
npx freellmapi doctor claude# shows whether claude actually reaches the gateway
You can also point other tools at it. Anything that takes an OpenAI compatible base URL can use http://localhost:3001/v1 with the unified key.
SOURCES AND DATES
Where these facts come from
FreeLLMAPI README and docs, read 25 Oct 2026. Router version 0.2.1, npm package freellmapi 0.8.2, catalogue snapshot v2026.10.24.
The 34 providers and 7.4 billion tokens a month come from the README and are its own estimate.
GitHub: the FreeLLMAPI repository, 41,065 stars, MIT licence, not archived, last push 24 Oct 2026 (UTC), checked 25 Oct 2026.
My own runs on 25 Oct 2026: the CLI dry run and the no providers reply, on Node 22 on Linux. I did not run the Docker route, the desktop app or a live reply from a provider.
That is the whole thing
Start with one tiny test through two providers. If replies are good enough for your side projects, keep it. If not, you lost ten minutes and paid nothing.
I'm Raj Singh. I've led ecommerce at two Fortune 100 companies, and at one of them I ran about $8M a year in ads for four years. I've worked with AI since 2019. Now my engineering team and I ship AI products, including debtops.ai. Each guide here is written from a real install, with the output dated.
