Remote OpenClaw Blog
Best Free Models for NanoClaw: Zero-Cost Local Agents
8 min read ·
The best free model for NanoClaw is gpt-oss:20b running locally through Ollama, which is Apache 2.0 licensed, fits on a machine with roughly 16 GB of memory, and pairs cleanly with NanoClaw's container-isolated execution model at zero API cost. If you cannot run a model locally, the strongest free cloud option is GPT-OSS 120B (openai/gpt-oss-120b:free) through OpenRouter's free tier, which has a 131K-token context window and no credit card requirement.
NanoClaw is a lightweight AI assistant that runs Claude agents inside isolated containers, documented at docs.nanoclaw.dev. Its documentation states that "Claude powers agents by default, with Codex, OpenCode, and Ollama as alternatives," which means Ollama is a first-class path for wiring free local models into a NanoClaw fork.
How NanoClaw Uses Models
NanoClaw runs each agent inside an isolated Apple Container or Docker sandbox, with a single Node.js process, SQLite message storage, and a per-group CLAUDE.md memory file. Because the codebase is deliberately small and fork-friendly, model choice is a code edit rather than a sprawling config surface, as described in the NanoClaw GitHub repository.
There are two practical zero-cost paths. The first is Ollama, which NanoClaw's docs list as a supported alternative to the default Claude backend. Running Ollama gives you an unlimited local model with no rate limits and no data leaving your device, which fits NanoClaw's security-centered, container-isolated design. The second is the Claude Agent SDK path: because that SDK honors the ANTHROPIC_BASE_URL environment variable, you can route it through OpenRouter's Anthropic-compatible endpoint and use free model IDs, per OpenRouter's Claude Code integration docs.
For a full primer on the framework itself, read what NanoClaw is, and for how it stacks up against nearby agents, see NanoClaw vs OpenClaw vs NemoClaw.
Best Free Models for NanoClaw, Ranked
The ranking below prioritizes models that are genuinely free, support reliable tool calling for agent work, and either run locally through Ollama or route through OpenRouter's free tier. All model IDs and context windows were verified against OpenRouter's free model collection as of July 2026.
1. GPT-OSS 20B (OpenAI)
GPT-OSS 20B is the best all-around free model for NanoClaw because it is Apache 2.0 licensed, runs locally on roughly 16 GB of memory, and delivers strong coding and agentic tool use. It has a 131K-token context window and is available both as an Ollama model (gpt-oss:20b) and free on OpenRouter (openai/gpt-oss-20b:free).
- Maker: OpenAI · Context: 131K · License: Apache 2.0
- How to use with NanoClaw: Pull it with Ollama and set Ollama as your agent provider so inference stays inside your container-isolated environment.
- Limitation: At 20B parameters it trails frontier models on the hardest multi-step reasoning chains.
2. GPT-OSS 120B (OpenAI)
GPT-OSS 120B is the strongest free model for NanoClaw when you offload inference to the cloud, matching much larger closed models on general reasoning while staying free on OpenRouter. It is a mixture-of-experts model with a 131K-token context, available as openai/gpt-oss-120b:free.
- Maker: OpenAI · Context: 131K · License: Apache 2.0
- How to use with NanoClaw: Route the Claude Agent SDK through OpenRouter with
ANTHROPIC_BASE_URL, or run it locally only if you have 60 GB or more of unified memory. - Limitation: Too large for most laptops to run locally, so you depend on the free tier's 20 requests per minute cap.
3. Qwen3 Coder (Alibaba Qwen)
Qwen3 Coder is the best free coding model for NanoClaw, offering a 1M-token context window and state-of-the-art open code generation. It is available as qwen/qwen3-coder:free on OpenRouter and as qwen3-coder through Ollama.
- Maker: Alibaba Qwen · Context: 1M · License: Apache 2.0
- How to use with NanoClaw: Use the Ollama build for coding-heavy repositories, or the OpenRouter free ID for large-context refactors.
- Limitation: Free availability on OpenRouter rotates and has dropped out for stretches, so keep a fallback configured.
4. Llama 3.3 70B Instruct (Meta)
Llama 3.3 70B is a dependable free generalist for NanoClaw with mature tool-calling behavior and broad ecosystem support. It has a 131K-token context and is offered as meta-llama/llama-3.3-70b-instruct:free and as llama3.3:70b in Ollama.
- Maker: Meta · Context: 131K · License: Llama 3.3 Community License
- How to use with NanoClaw: Best via OpenRouter free unless you have roughly 40 GB of memory for the local Ollama build.
- Limitation: The 70B local build is heavy, and the free OpenRouter route is deprioritized during peak traffic.
5. Gemma 4 31B (Google)
Gemma 4 31B is a strong free option for NanoClaw workflows that need a large context, with a 256K-token window and high reasoning quality per parameter. It is available on OpenRouter as google/gemma-4-31b-it:free.
- Maker: Google · Context: 256K · License: Gemma terms of use
- How to use with NanoClaw: Route through OpenRouter free for long-document or long-conversation groups.
- Limitation: Tool-calling reliability is less battle-tested than GPT-OSS and Llama for multi-tool agent chains.
Free Model Comparison Table
The table below compares the top free NanoClaw models on maker, context window, best use, and how each connects to NanoClaw.
| Model | Maker | Context | Best For | NanoClaw Path |
|---|---|---|---|---|
| GPT-OSS 20B | OpenAI | 131K | Local default, coding, agents | Ollama (local) |
| GPT-OSS 120B | OpenAI | 131K | Strongest free reasoning | OpenRouter free |
| Qwen3 Coder | Alibaba Qwen | 1M | Coding, huge context | Ollama or OpenRouter free |
| Llama 3.3 70B | Meta | 131K | General-purpose, tool calling | OpenRouter free (or Ollama) |
| Gemma 4 31B | 256K | Long-context reasoning | OpenRouter free |
For a broader zero-cost menu that applies across agent stacks, see our OpenRouter free models guide and the wider best free models for OpenClaw rundown.
How to Wire a Free Model Into NanoClaw
NanoClaw supports two zero-cost setups: a local Ollama model or the Claude Agent SDK routed through OpenRouter. The local path is the most private and has no rate limits.
Option A: Local Free Model via Ollama
Ollama is listed in NanoClaw's docs as an alternative agent provider, so a local model runs inside your own environment with no API cost and no usage caps.
# Install Ollama and pull the recommended free model
curl -fsSL https://ollama.com/install.sh | sh
ollama pull gpt-oss:20b
# Serve with enough context for agent tool loops
OLLAMA_CONTEXT_LENGTH=64000 ollama serve
# In your NanoClaw fork, select Ollama as the agent provider
# and point it at the local endpoint: http://localhost:11434
Option B: OpenRouter Free Models via the Claude Agent SDK
Because NanoClaw runs Claude agents by default and the Claude Agent SDK honors ANTHROPIC_BASE_URL, you can route agents through OpenRouter and select a free model ID.
# Point the Claude Agent SDK at OpenRouter's Anthropic-compatible endpoint
export ANTHROPIC_BASE_URL="https://openrouter.ai/api"
export ANTHROPIC_AUTH_TOKEN="your-openrouter-key"
export ANTHROPIC_API_KEY="" # explicitly blank to avoid conflicts
# Override the model tiers with a free OpenRouter model
export ANTHROPIC_DEFAULT_SONNET_MODEL="openai/gpt-oss-120b:free"
OpenRouter warns that the Claude Agent SDK is tuned for Anthropic models, so treat open models over this route as best-effort for complex multi-tool chains and keep a fallback ready.
Limitations and Tradeoffs
Free models make NanoClaw genuinely zero-cost, but they come with real constraints. Local Ollama models need suitable hardware: an 8B model wants about 8 GB of memory, while the recommended gpt-oss:20b wants roughly 16 GB and a 70B model needs around 40 GB. If you do not already own that hardware, buying it to avoid a few dollars per month in API costs rarely makes financial sense.
The OpenRouter free tier is capped at 20 requests per minute with roughly 200 requests per day per model, and free requests are deprioritized during peak traffic. That is fine for a single-user NanoClaw assistant but too tight for always-on or multi-group deployments. Free model availability also rotates: DeepSeek's free variants, for example, lost their free status on OpenRouter in mid-2026, so any free ID should be treated as subject to change.
When not to use free models with NanoClaw: client-facing automation, unattended overnight runs, or complex multi-step tool chains where a failure has consequences. For those, a cheap paid model is more reliable. See best cheap models for OpenClaw for the next tier up from free.
Related Guides
- What Is NanoClaw?
- NanoClaw vs OpenClaw vs NemoClaw
- OpenRouter Free Models Guide
- Best Free Models for OpenClaw
Go deeper
The operator playbooks
Production-ready PDF guides for OpenClaw and Hermes Agent — $19.99 each.
Skills for this topic
Browse all skills →Frequently Asked Questions
What is the best free model for NanoClaw?
The best free model for NanoClaw is GPT-OSS 20B ( gpt-oss:20b ) running locally through Ollama. It is Apache 2.0 licensed, fits on roughly 16 GB of memory, has a 131K-token context, and runs inside NanoClaw's container isolation with no API cost or rate limits. If your hardware is too small, GPT-OSS 120B on OpenRouter's free tier is the strongest
Can NanoClaw run local models for free?
Yes. NanoClaw's documentation lists Ollama as an alternative agent provider to the default Claude backend. You install Ollama, pull a model such as gpt-oss:20b or qwen3-coder , and point your NanoClaw fork at the local endpoint. There is no API key, no billing, and no usage cap beyond your own hardware.
How do I use OpenRouter free models with NanoClaw?
NanoClaw runs on the Claude Agent SDK, which honors the ANTHROPIC_BASE_URL environment variable. Set it to https://openrouter.ai/api , provide your OpenRouter key as ANTHROPIC_AUTH_TOKEN , blank out ANTHROPIC_API_KEY , and override the model tier with a free ID like openai/gpt-oss-120b:free . OpenRouter notes this route is tuned for Anthropic models, so open models are best-effort for complex tool use.
What context window do free NanoClaw models offer?
As of July 2026, GPT-OSS 20B and 120B offer 131K tokens, Llama 3.3 70B offers 131K, Gemma 4 31B offers 256K, and Qwen3 Coder offers up to 1M tokens on OpenRouter's free tier. For agent tool loops, configure at least 64K of context so NanoClaw does not lose track of instructions and tool state.
Are free models good enough for production NanoClaw agents?
For personal assistants and light workloads, yes. For client-facing work, unattended runs, or long multi-tool chains, free models fail more often than paid ones, and the OpenRouter free tier's 20 requests per minute cap can throttle heavier use. A cheap paid model is the safer choice for reliability-critical NanoClaw deployments.