That AUTH Error is lying. Your API key is almost never the problem — the session is stale, or your model can’t do the job. Two curls tell you which.
Free Claude Code = a middleman app on your PC. Claude Code phones Anthropic; this answers the call itself and asks NVIDIA’s free models instead. Same terminal, $0 instead of $240/year.
Straight up: this does not hand you Claude. You keep the Claude Code tool — the thinking comes from someone else’s model.
🧠 What this actually is, in 30 seconds — no background needed
Claude Code is Anthropic’s coding assistant that lives in your terminal — it reads your files, writes code, runs commands, fixes its own mistakes. It normally costs $20/month (Pro) up to $200/month (Max).
Free Claude Code (FCC) is a small program you run on your own machine. Claude Code sends every request to Anthropic’s servers — FCC quietly stands in the middle, catches those requests, and forwards them to a different AI company that has a free tier. Claude Code never notices the swap.
├─ 
You keep: the terminal, the file editing, the agent behaviour, /commands, everything
├─ 
You swap: the brain — NVIDIA, DeepSeek, Gemini, Groq, Kimi, MiniMax, Z.ai, or a model on your own PC (Ollama / LM Studio / llama.cpp). 28 providers, switch anytime
└─ 
You don’t get: actual Claude models. Nobody gives those away
MIT licensed · 42,000+ stars · shipping updates the same week. It also launches Codex and Pi the same way (fcc-codex, fcc-pi).
⚡ The fix — changing the model mid-session does nothing
A live Claude session stays glued to whatever model it booted with. Pick a new one from the menu and it keeps hammering the old, dead one — throwing errors that look like auth failures.
# 1. exit Claude completely (not a new tab — fully out)
# 2. restart the proxy
fcc-server restart
# 3. brand-new session
fcc-claude
First message to prove it’s alive:
Reply with exactly: FCC is working
Second cause, same symptom — and it’s FCC’s most-reported bug. If you ever logged into claude.ai from the terminal, that saved login outranks your proxy token. You get:
Auth conflict: Both a token (ANTHROPIC_AUTH_TOKEN) and an API key (/login managed key) are set.
API Error: 401 {"detail":"Invalid API key"}
Your key is fine. One command clears it:
claude /logout
…then relaunch with fcc-claude. 
Also check cat ~/.claude/settings.json | jq .env — a value saved in that file silently beats anything you export in your shell.
🩺 Two curls that split the blame in 10 seconds — key vs proxy
① Is my NVIDIA key good?
curl -s -o /dev/null -w "%{http_code}\n" \
-H "Authorization: Bearer $NVIDIA_API_KEY" \
https://integrate.api.nvidia.com/v1/models
200 → key is fine, account is fine. Stop making new keys. · 401 → no key sent · 403 → key rejected or that model isn’t granted to you — 403 alone can’t tell you which.
② Is the proxy actually answering?
curl -X POST "http://localhost:8082/v1/messages" \
-H "Authorization: Bearer freecc" -H "anthropic-version: 2023-06-01" \
-H "content-type: application/json" \
-d '{"model":"claude-sonnet-4-6","max_tokens":1,"messages":[{"role":"user","content":"."}]}'
Anything starting {"id":"msg_ → working. Even an unknown-model error counts as a pass — it authenticated you before rejecting the name.
Both curls green but Claude Code still 401s? Then it’s the stale session above, 100%. ANTHROPIC_BASE_URL must be the root — http://localhost:8082, no /v1. Claude Code adds that itself.
🧨 The model trap — anything named "coder" quietly breaks it
Claude Code needs tool calling — the model’s ability to actually run commands and edit files, not just talk about them. Without it you get the worst kind of failure: no error at all. Claude describes what it’s about to do, prints tool JSON as plain text, and never touches a single file.
Trap: models with coder, codellama, starcoder, codestral or granite-code in the name are mostly 2023-era builds from before tool calling existed. They’re the obvious-looking pick in the catalog and they’re the wrong one — deepseek-coder-6.7b-instruct is exactly this.
Picks that work:
├─ nvidia/nemotron-3-super-120b-a12b — FCC’s default, tool calling confirmed (120B total brain, 12B used per answer — fast for its size)
├─ minimaxai/minimax-m2.7 — the one with a confirmed working Claude-Code-on-NIM report; switch to it the moment tools stop firing
└─ 
Full live list: integrate.api.nvidia.com/v1/models — 118 models, no key needed to look
That list carries zero capability info — presence proves nothing. Smoke-test any model before you trust it. Context limits: NVIDIA’s docs claim up to 1M, sibling models measured ~131K — test your own ceiling.
🚀 Full setup from zero — install, run, Linux needs two terminals
curl -fsSL "https://raw.githubusercontent.com/Alishahryar1/free-claude-code/main/scripts/install.sh" | sh
Same command updates it later. Add --dry-run first to see what it does without doing it.
It also installs Claude Code and Codex, and offers Pi — not a single-package install. It pulls its own Python 3.14 through uv; on Ubuntu that’s expected, not a bug. It refuses to run while any fcc-* process is alive.
fcc-server # terminal 1 — leave it open, this IS the proxy
fcc-claude # terminal 2 — your Claude Code session
No tray icon on Linux — that’s Windows/macOS only, so two terminals is normal, not a broken install.
Paste your NVIDIA key into the admin panel: http://127.0.0.1:8082/admin (your PC only, nobody else can reach it).
Get the key free at build.nvidia.com.
Real commands are fcc-server fcc-claude fcc-codex fcc-pi. (fcc-init and free-claude-code aren’t commands — you’ll get command-not-found.)
⚠️ Read before you start — phone gate, blocked countries, rate limit, terms
├─ 
No credit card — but a phone number is mandatory. NVIDIA sends an SMS code before it’ll issue a key.
├─ 
Country blocks on that SMS step: TR, IR, SY, CU, KP, RU, BY. If you’re in one, you hit a wall at the key step — use a different provider in FCC instead (OpenRouter, Gemini, Groq, Cerebras all have free tiers).
├─ 
~40 requests per minute is the real free-tier ceiling. No official way to raise it. Old “1,000 credits” numbers floating around are years stale.
├─ 
Terms: NVIDIA keeps what you send and what comes back to improve their models, and the free API is for testing and evaluation — not production. So: personal projects, learning, messing around · client code, anything under NDA
└─ 
Never overwrite ~/.claude.json with a fresh file — merge into it, or you wipe your whole Claude Code config.
💡 Five moments this actually saves you
├─ 
You want to try an agent that edits your own files, but $20/month to find out if you even like it is a hard no.
├─ 
You hit your plan’s limit at 11pm mid-task — flip to fcc-claude and keep going on the free backend instead of stopping.
├─ 
You’re on your 4th regenerated API key at 2am. All four were valid. The curl would have said so in ten seconds.
├─ 
Claude confidently explains the fix, “writes” the file… and nothing changes on disk. That’s the tool-calling trap, not you doing it wrong.
└─ 
You’ve got a decent GPU sitting idle — same tool, pointed at Ollama or LM Studio, fully offline, no key at all.
Fast version: curl the key → 200 → claude /logout → quit → fcc-server restart → new session. And if tools never fire, it’s the model, not you.
A key that answers 200 has nothing left to confess — everything after that is the client’s fault, not your account’s.
!