You can run Claude Code, Anthropic's agentic coding CLI, for $0/month in 2026 by routing it through OpenRouter's free tier instead of paying for Anthropic's API directly. The whole setup is three environment variables — no proxy, no Docker, no credit card. Claude Code became effectively open source after a March 31, 2026 source-code leak exposed all 512,000 lines, and OpenRouter's "Anthropic Skin" speaks Claude Code's native protocol, so Thinking blocks, tool use, and streaming all pass through untouched. Free models are weaker than paid Claude, but for builders, small-business owners, and anyone prototyping, the cost-to-capability ratio is hard to beat.
Last verified: 2026-07-29 · Claude Code v2.1.220 · OpenRouter free tier (28+ free models) · Pricing/limits change often — re-check before relying on this for production.
What changed: why free Claude Code is possible now
Two things had to happen for $0 Claude Code to work, and both landed in 2026.
First, Claude Code got an effectively open-source blueprint. On March 31, 2026, security researcher Chaofan Shou discovered that version 2.1.88 of Anthropic's @anthropic-ai/claude-code npm package contained a 59.8 MB source map file (cli.js.map) that exposed the entire 512,000-line, 1,900-file TypeScript codebase (AkitaOnRails analysis). Within hours, the code was mirrored to GitHub and forked 41,500+ times. Anthropic pulled the file, but the cat was out of the bag — the community had the full architecture, the feature flags, the telemetry hooks, and the tool-use protocol. Today, the official repo at github.com/anthropics/claude-code carries 140,000+ stars and 7.9 million weekly downloads (npm).
Second, OpenRouter built the "Anthropic Skin" — a native endpoint compatible with the Anthropic Messages API. That means Claude Code speaks its own protocol straight to OpenRouter with no translation proxy in between. Thinking blocks, native tool use, streaming, and multi-turn context all work as they do against Anthropic directly (OpenRouter blog). Before this, you needed a local proxy like claude-code-router running a Node.js process on a local port — a whole tier of infrastructure that used to break.
How to set up free Claude Code with OpenRouter (step by step)
The setup is three environment variables on top of a working Claude Code install. No proxy, no Docker, no local port to keep alive.
Step 1: Install Claude Code
Claude Code is an npm package. You need Node.js 18+ installed first.
# Install Node.js (if you don't have it) — LTS from nodejs.org
# Then install Claude Code globally
npm install -g @anthropic-ai/claude-code
# Verify
claude --version
Weekly downloads are pushing 7.9 million as of July 2026 (npm), so you are running what most developers run. Current version is v2.1.220.
Step 2: Create an OpenRouter account and API key
- Go to openrouter.ai and sign up with email or GitHub. No credit card required for the free tier.
- Open Keys from the dashboard.
- Click Create Key and copy the result (it starts with
sk-or-v1-...).
OpenRouter's free tier gives you 20 requests per minute and 50 requests per day against 28+ free models — rising to 1,000 requests per day once you add $10 in lifetime credits (OpenRouter free tier guide). That $10 is a one-time purchase that never expires; it is not a subscription.
Step 3: Point Claude Code at OpenRouter
Add three environment variables to your shell profile (~/.zshrc on macOS, ~/.bashrc on Linux):
export ANTHROPIC_BASE_URL="https://openrouter.ai/api"
export ANTHROPIC_AUTH_TOKEN="sk-or-v1-your-key-here"
export ANTHROPIC_API_KEY=""
Source the file, then confirm the switch inside Claude Code with /status:
source ~/.zshrc
cd your-project
claude
> /status
# Should show: Anthropic base URL: https://openrouter.ai/api
Your requests should appear in OpenRouter's Activity dashboard within seconds. That is the entire setup. The official integration guide is at the OpenRouter docs (Claude Code integration).
Step 4: Pick a free model
Free model IDs end in :free. These are reliable coding-friendly picks as of July 2026:
| Model ID | Context | Best for | Source |
|---|---|---|---|
qwen/qwen-2.5-coder-32b-instruct:free |
128K | General coding, refactors | OpenRouter models |
mistralai/devstral-2512:free |
128K | Code completion, agentic loops | OpenRouter free roster |
deepseek/deepseek-r1:free |
128K | Reasoning, math, multi-step plans | OpenRouter free roster |
meta-llama/llama-3.3-70b-instruct:free |
128K | General-purpose, well-supported | OpenRouter free roster |
google/gemini-flash:free |
1M | Multimodal, huge context | OpenRouter free roster |
The roster shifts month to month as upstream providers add or pull free hosting. Browse the live list at openrouter.ai/models?max_price=0. If you already grapple with token limits on paid Claude, a separate guide walks through the 15 habits that actually cut waste regardless of which provider you use (How to never hit a Claude token limit again).
Step 5: Configure auto-fallback so your session never dies
The single biggest quality-of-life win is OpenRouter's automatic fallback. If your primary model's provider is down, rate-limited, or refuses the reply, OpenRouter walks a fallback list you define and tries the next model — automatically, with no code change on your side (OpenRouter model fallbacks docs).
Pass the fallback chain in the models array of your API request. For Claude Code, you can configure this in your project's settings so every request tries a primary model, then quietly swaps:
- Primary:
qwen/qwen-2.5-coder-32b-instruct:free - Fallback 1:
mistralai/devstral-2512:free - Fallback 2:
deepseek/deepseek-r1:free - Fallback 3:
meta-llama/llama-3.3-70b-instruct:free
The moment one model's free tokens run out or its provider rate-limits you, the next one picks up. You do not touch a thing. This is what makes a free setup viable for sessions longer than a few minutes — a real concern we cover in detail in the AI coding agent divergence guide, which digs into why agents get stuck polishing one answer and how to break them out.
Step 6 (optional): Use the ccr router for full control
If you want finer-grained routing — different models for background tasks vs. thinking vs. long-context work — the open-source claude-code-router (ccr) project wraps the same idea in a local config file. You define a Router block with default, background, think, and longContext model slots, then run ccr start and ccr code instead of claude directly (GitHub setup guide). This adds a local Node.js process but gives you per-task-class model routing and a longContextThreshold so big files automatically go to a higher-context model.
What it actually costs: free vs. paid Claude Code
The math is the whole point. Here is the honest comparison.
| Setup | Monthly cost | What you get | Best for |
|---|---|---|---|
| Claude Code + Anthropic API (Opus) | ~$50–$200+ (usage) | Frontier model, full context, no rate limits | Production sessions, complex refactors |
| Claude Max subscription | $100–$200 (flat) | Flat-rate Claude, higher caps | Heavy daily users who want predictability |
| Claude Code + OpenRouter free tier | $0 | 28+ free models, 50 req/day, auto-fallback | Prototyping, learning, hobby projects |
| Claude Code + OpenRouter + $10 credits | $0 (one-time $10) | 1,000 req/day, all free models | Builders shipping side projects |
| Claude Code + OpenRouter BYOK | $0 routing (1M req/mo) + your provider costs | Use your own OpenAI/Anthropic/Google keys via OpenRouter's gateway | Teams with existing provider credits |
Cost figures are drawn from OpenRouter's published free tier guide and Anthropic's published pricing pages; token consumption on Opus can exceed $200/month for heavy users (OpenRouter blog, OpenRouter free tier). The $10 credit is a one-time purchase — it permanently raises your daily cap to 1,000 and never expires. It is not a recurring bill.
Is free Claude Code good enough for real work?
Honest answer: for first drafts, yes. For production sessions on hard problems, not always.
The free models on OpenRouter are strong open-weights — Qwen3 Coder 480B, DeepSeek R1, Llama 3.3 70B, Gemini Flash. They are competent at code explanation, test generation, standard refactors, and multi-file edits. They are not Claude Opus 5 or GPT-5.6. On complex reasoning, long-context refactors, and subtle bug-hunting, the gap is real.
The practical pattern most builders settle on is hybrid: run free models for the 80% of work that is mechanical — boilerplate, tests, docs, scaffolding — and keep a small paid Anthropic budget ($10–20/month in OpenRouter credits) for the 20% that needs frontier intelligence. Because OpenRouter is a single gateway, you switch models with one config change, not a new account. If you want the deeper framework for deciding when a single agent is enough vs. when you need a paid multi-agent setup, see our single-agent vs. multi-agent decision framework.
Which coding tools work with OpenRouter's free tier?
OpenRouter is model-agnostic — it routes any OpenAI- or Anthropic-compatible client. That means the same free "brain" works behind several coding tools, so you can pick the interface you like without changing your model setup.
| Tool | License | Open source? | Works with OpenRouter free? | Source |
|---|---|---|---|---|
| Claude Code | Proprietary (source leaked Mar 2026) | Effectively yes (512K-line source public) | Yes — via Anthropic Skin, 3 env vars | github.com/anthropics/claude-code |
| Cline | Apache 2.0 | Yes (58K+ stars) | Yes — BYOK, multi-model | cline.bot |
| Aider | Open source | Yes | Yes — any LLM API | Aider docs |
| Cursor | Proprietary | No | Bring your own model | cursor.com |
| GitHub Copilot | Proprietary | No | Multi-model (Microsoft default) | GitHub |
| OpenAI Codex | Proprietary | No | Consumption per task | OpenAI |
Cline is worth a separate mention: it is an open-source VS Code extension with 58,000+ GitHub stars and Apache 2.0 licensing, and it brings your own API key by default (Cline). If you prefer an IDE-resident agent over a terminal one, Cline + OpenRouter free is the equivalent free path. The choice between Claude Code (terminal) and Cline (IDE) is mostly about where you like to work, not about capability.
Can you run multiple free agents at the same time?
Yes — and this is where the setup compounds. Because OpenRouter keys are not tied to a single session, you can run several agents in parallel, each pointed at the same key, each building a different thing.
A realistic small-business pattern from the research:
- Agent 1 (Claude Code + Qwen Coder free): builds a landing page
- Agent 2 (Claude Code + DeepSeek R1 free): writes email follow-ups
- Agent 3 (Cline + Llama 3.3 70B free): cleans up a customer list
All three run in the background, all on free models, all behind the same OpenRouter key with auto-fallback. You check back later and the work is done. The constraint is the 20-requests-per-minute cap across all sessions combined — so parallel agents share that bucket. If you hit the wall, the $10 credit bump to 1,000 requests/day removes the ceiling for almost everyone.
This is also the pattern behind a full "AI agent OS" — a local-first layer that schedules agents, shares memory, and routes tools. We wrote a full guide on how to build an AI agent OS in 2026 that covers the architecture, the memory piece, and which parts you should and should not automate.
What this means for you
If you are a small-business owner or solo builder: Start with one agent and one simple job — a landing page, an email draft, a lead form. Watch it run on a free model. The skill that matters now is not coding — it is knowing what to ask for. If you know what a good landing page looks like and what a good email says, you already have the hard part. The free AI handles the rest. The same logic applies to broader AI for small business workflows where the bottleneck is specification, not execution.
If you are a developer: Use the free tier for prototyping, learning Claude Code's workflow, and the mechanical 80% of daily work. Put $10 into OpenRouter once to lift the daily cap, keep auto-fallback on, and switch to paid Anthropic only for sessions that genuinely need frontier reasoning. Your monthly bill drops from $50–$200 to single digits for most workloads.
If you are evaluating a stack for a team: OpenRouter's BYOK program lets you bring existing OpenAI/Anthropic/Google credits in and gives you 1 million free routing requests per month per account. That is roughly 32,000 requests/day — plenty for most teams. You get unified billing, per-key budget caps, and an Activity dashboard that shows cost per session. The tradeoff is the 5% routing fee above 1M requests and a small credit-purchase fee (5.5%, $0.80 minimum) (OpenRouter blog).
FAQ
Q: Is Claude Code actually free?
A: The CLI tool itself has always been free to install. What costs money is the model it talks to. Routing Claude Code through OpenRouter's free tier — 28+ free models with :free IDs — means the inference is also $0, subject to rate limits (20 RPM, 50 requests/day on the unfunded tier, 1,000/day after $10 in lifetime credits).
Q: Did Anthropic open-source Claude Code?
A: Not officially. On March 31, 2026, a misconfigured npm package (v2.1.88) accidentally bundled a 59.8 MB source map containing all 512,000 lines of Claude Code TypeScript. Security researcher Chaofan Shou reported it; the code was mirrored to GitHub before Anthropic pulled the file. The official repo is at github.com/anthropics/claude-code with 140K+ stars. License terms say "SEE LICENSE IN README.md," so treat the source as readable-but-not-OSI-licensed rather than truly open source.
Q: What is OpenRouter's free tier limit in 2026?
A: 20 requests per minute, 50 requests per day on the unfunded free tier, and 1,000 requests per day once you have added $10 in lifetime credits. Token-per-minute caps depend on the underlying provider. Free model IDs must end in :free — deepseek/deepseek-r1 without the suffix is paid passthrough, not the free variant (OpenRouter free tier guide).
Q: Does Claude Code work with OpenRouter without a proxy?
A: Yes. OpenRouter operates an "Anthropic Skin" — an endpoint compatible with the Anthropic Messages API — so Claude Code speaks its native protocol directly to OpenRouter. Thinking blocks, tool use, streaming, and multi-turn context all pass through. You set three environment variables (ANTHROPIC_BASE_URL, ANTHROPIC_AUTH_TOKEN, ANTHROPIC_API_KEY) and skip the local proxy entirely (OpenRouter blog).
Q: What happens when a free model runs out of tokens or gets rate-limited?
A: OpenRouter's auto-fallback walks a models array you define and tries the next model automatically — no code change, no manual swap. You configure a fallback chain (primary → fallback 1 → fallback 2) and OpenRouter handles the rest. Provider-level failover is real but cannot route around OpenRouter itself going down, so plan for both (OpenRouter model routing).
Q: Are free models on OpenRouter good enough for production coding?
A: For prototyping, learning, boilerplate, tests, and standard refactors — yes. For complex reasoning, long-context refactors, and subtle bugs — generally no. The practical pattern is hybrid: free models for the 80% of mechanical work, a small paid Anthropic budget for the 20% that needs frontier intelligence. OpenRouter lets you switch models with one config change.
Discussion
0 comments