NanoClaw costs $0. It is MIT-licensed, self-hosted, has no user accounts and no paid tier, and its documentation site publishes no pricing page at all — so "NanoClaw pricing" means the model bill from whichever provider you attach, plus a Docker Desktop subscription if your company is above Docker's free-use thresholds. The number worth planning is what the default runtime bills, because NanoClaw ships pointed at Claude; Codex is the one alternative setup can install inline, and the rest are add-ons you install afterwards.
Two disambiguations first, because the results page mixes them in. Clawland-AI/nanoclaw is an unrelated Python agent for single-board computers, and its hardware costs are not this product's. A widely-shared third-party headline also attaches a dollar figure to the words "NanoClaw PR factory"; whatever that figure measures, it is not a price NanoClaw charges, because NanoClaw charges nothing. This page is about nanocoai/nanoclaw — MIT, not archived, 30,814 stars, created January 31, 2026, last pushed September 19, 2026 (GitHub API, September 21, 2026). The older qwibitai/nanoclaw address now returns a 301 to that same repository, so a write-up or directory entry still pointing there is not describing a second project. The install NanoClaw documents is bash nanoclaw.sh from that checkout, so treat a SourceForge listing as a mirror rather than the project's own channel.
What the software costs, and what actually bills you
NanoClaw's homepage FAQ answers the money question in one line: it is free and open source under the MIT license, and "your agent provider may charge for model usage." The documentation index lists 56 pages and none of them is a pricing, billing or plans page. The bill lives entirely on the other side of the container.
| Line item | What it costs | Where that comes from |
|---|---|---|
| NanoClaw itself | $0, MIT, no accounts | Repository license, and the README's "NanoClaw has no user accounts" |
| Free NanoClaw account | $0 and optional — two perks only | Local image builds are the default and need no account, per the hardened-image doc |
| OneCLI credential vault | $0 self-hosted (Apache-2.0); its ee/ features need a subscription for production use | OneCLI repository; NanoClaw installs a local instance by default |
| Docker | Docker Desktop Personal $0; Pro $11, Team $16 and Business $24 per user per month billed monthly, with Pro $9 and Team $15 annually | Docker pricing; free use is capped by the licensing terms at personal, education, non-commercial open source, and businesses under both 250 employees and $10 million revenue |
| Model tokens | Your provider's per-token rate | Your provider's own billing |
All checked September 21, 2026. The Docker line is the one people miss: Docker Engine on Linux does not trigger that clause, but a commercial macOS or Windows install above those thresholds does. Before any of it, the README requirements are macOS or Linux (Windows via WSL2), Node.js 22+ and pnpm 10+, Docker, and Claude Code itself — which is a prerequisite for /customize, /debug, setup error recovery and every /add-<channel> skill, not only for the default provider. Current release is v2.3.0, published August 24, 2026, and package.json on main agrees.
The default runtime is Claude, and there are three ways to pay for it
This is the distinction most provider lists blur. Out of the box NanoClaw runs Claude Code inside the agent container through Anthropic's Claude Agent SDK — container/cli-tools.json pins @anthropic-ai/claude-code at 2.1.257. NanoClaw's credentials guide describes three ways the wizard connects Claude — a Claude subscription, an API key or an OAuth token you already hold, or an Anthropic-compatible endpoint — and states that all three end up as a vault secret rather than an environment variable, with the subscription path running claude setup-token and storing the resulting OAuth token. Vendor prices below are the amounts those pages showed on the date given; confirm the figure in the vendor's own checkout before you commit to it.
On the subscription option, two official statements sit side by side and this page will not reconcile them for you. NanoClaw's wizard labels subscription sign-in "recommended if you have Pro or Max", and Claude's pricing page lists Pro at $20 per month billed monthly, or $17 per month on the annual plan at $200 billed up front, and Max from $100 per month. Anthropic's Agent SDK overview separately states that, unless previously approved, it does not allow third-party developers to offer claude.ai login or rate limits for their products, including agents built on the Agent SDK. Whether NanoClaw holds such an approval is not something any public source here could establish, and neither is the usage quota a subscription gives an agent — Anthropic publishes no token figure for Pro or Max.
NanoClaw custom API: the one documented route
An Anthropic-compatible endpoint is a first-class setup path, and it is narrower than the README suggests. Setting both variables below makes setup create a OneCLI secret with header injection keyed to your hostname, write only the base URL into .env, and register the custom-endpoint provider config.
# NanoClaw's credentials guide: set BOTH before running setup.
# Only one of the two is not enough — setup takes this branch only when
# no Anthropic secret exists yet, and the URL is an origin with no /v1.
export NANOCLAW_ANTHROPIC_BASE_URL=https://api.kunavo.com
export NANOCLAW_ANTHROPIC_AUTH_TOKEN=<your key>
# The same two values exist as setup flags:
# --anthropic-base-url --anthropic-auth-tokenFour constraints follow from how that is built. Both variables are required — setup's branch tests for the pair. The branch only runs when no Anthropic secret exists yet, so this is a first-install path, not a way to re-point a machine that already has Claude connected. The secret binds to the hostname parsed out of your URL, so changing domains means rebuilding it. And the URL must be an origin: Kunavo's own ANTHROPIC_BASE_URL reference spells out why https://api.kunavo.com works and https://api.kunavo.com/v1 produces requests to /v1/v1/messages.
Inside the container your token never appears. NanoClaw's Claude provider injects the base URL plus ANTHROPIC_AUTH_TOKEN=placeholder, and the vault's gateway rewrites the bearer on the wire — which also means the injection depends on the container using its configured HTTPS_PROXY. Anything that bypasses that proxy bypasses the credential too. In v2.3.0 the driver refuses a container spec that carries a credential value in contributed environment at all, so hand-editing a real key into a generated container.json is denied by policy rather than merely discouraged.
Two official sources disagree about the plain .env version, and both are worth knowing. The README presents ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN in .env "for one-off experiments", while the environment-variable reference says ANTHROPIC_BASE_URL is only active once setup has registered the custom-endpoint provider config. On current main the token half never reaches a request: the Claude provider file linked above asks .env for ANTHROPIC_BASE_URL only, and sets ANTHROPIC_AUTH_TOKEN to the literal string placeholder. Treat the README snippet as an experiment, exactly as it is labelled, and the setup route above as the configuration you would run.
Default versus extended: what each route actually installs
| Route | How you get it | Where credentials live | What it adds or costs |
|---|---|---|---|
| Claude (default) | Ships in the box | OneCLI vault, host pattern api.anthropic.com | Anthropic API key, or a Pro/Max subscription token |
| Anthropic-compatible endpoint | Two setup variables, first install only | Vault, keyed to your hostname | Your gateway's per-token rate; no NanoClaw fee |
| OpenCode | /add-opencode | Vault, with a matching host pattern — never .env | Routes via OpenCode's own config: OpenRouter, DeepSeek, OpenAI, Google, Anthropic or OpenCode Zen. On a pinned hardened-runtime install it edits the agent-runner lockfile, so the rebuild drops to a local build |
| Codex | Setup's runtime picker installs it inline, or /add-codex later | Vault, host pattern api.openai.com / chatgpt.com; the host strips OPENAI_API_KEY | ChatGPT subscription or OpenAI API key; ChatGPT lists Go $8, Plus $20, Pro from $100, Business $20 per user per month billed annually |
| Ollama (local) | /add-ollama-provider — see caveat | Not applicable; per NanoClaw's Ollama notes the route sets NO_PROXY so requests skip the credential-injecting proxy | No hosted-model bill; hardware and operations instead |
Sources checked September 21, 2026: NanoClaw's providers guide, its /add-opencode and /add-codex skill files, docs/ollama.md, and ChatGPT's pricing page (the old Codex pricing URL now redirects there). The first column follows the same providers guide, which says setup's runtime picker offers Codex as an inline install while OpenCode is added after setup with /add-opencode. The Ollama caveat is the vendor's own: the providers guide states the skill remains in the repository but targets an older configuration path, that those legacy edits are "not a supported drop-in switch on current main", and that the integration needs porting and testing first — while the README still lists it as a supported path. Where the two disagree, the newer and more specific statement is the providers guide. OpenRouter reaches NanoClaw through the OpenCode backend, not through an Anthropic base URL; OpenRouter's FAQ charges 5.5% with a $0.80 minimum on card credit purchases and 5% on crypto, with no markup on inference itself.
A worked NanoClaw API cost estimate
This is token arithmetic on stated assumptions, not a measured task cost and not a bill ceiling. Assume one agent group handling 40 turns a day, each sending 30,000 input tokens and returning 1,500 output tokens, with nothing served from cache. Rates are live Kunavo catalog prices per million tokens.
| Model | Input / output per 1M | Estimated day | Estimated 30-day month |
|---|---|---|---|
| Claude Haiku 4.5 | $0.40 / $2.00 | $0.60 | $18.00 |
| Claude Sonnet 4.6 | $1.20 / $6.00 | $1.80 | $54.00 |
| Claude Sonnet 5 | $2.00 / $10.00 | $3.00 | $90.00 |
| Claude Opus 5 | $2.00 / $10.00 | $3.00 | $90.00 |
Three readings. First, Claude Sonnet 5 and Claude Opus 5 carry the same catalog rate here, so on this shelf that step is a capability decision rather than a price one — against the rate Anthropic publishes for that same Opus 5 model id, $5.00 and $25.00 per million, the identical month models out at $225.00. Second, caching is the bigger lever than the model row, because that same page puts a cache read at 0.1x base input and a five-minute cache write at 1.25x: if 70% of the input above arrived as cache reads at the catalog's cached-input rate of $0.20 per million, the same Claude Opus 5 month models at $44.64 instead of $90.00. Third, scale the turn count to your own channels before treating any of this as a budget — a NanoClaw agent wired to a busy group chat and one wired to a single person are not the same workload.
Kunavo's catalog amount is a billing floor rather than a cap: when the upstream reports its charge, the bill is the greater of catalog cost and upstream cost times the applicable markup. The minimum top-up is $10 in prepaid credit — a funding minimum, not a task fee or a subscription. See billing details and usage reporting.
Minimum verification, and what extending costs you later
The cheapest useful check is small: setup/index.ts registers a verify step, and the /add-opencode skill's own acceptance is to send one message, confirm a reply, then send a second to confirm the session continues. For a custom endpoint, docker exec <container> env | grep ANTHROPIC should show your base URL next to a placeholder token — if the real key is visible there, something is wrong with the path, not right with it.
Then price the maintenance, because extended providers are pinned software. NanoClaw's changelog records that the previous Codex pin defaulted to a model OpenAI retired from Codex on August 31, 2026, and that re-running /add-codex does not re-pin an existing install — the current pin is @openai/codex 0.146.0. OpenCode's pin is itself a live source conflict: the repository's skill file says both pins must remain exactly 1.18.25, while the published providers guide still says 1.4.17. The repository is the more specific source, and OpenCode has separately moved organizations — sst/opencode now 301s to anomalyco/opencode, latest release v1.18.31 on September 14, 2026. Installs on OneCLI 1.41.0 also need manual reauthentication when an OpenCode ChatGPT login expires.
Capability changes when you leave the Claude provider, per the providers guide: OpenCode and Codex both declare no native slash commands, both queue follow-up messages instead of accepting mid-turn input, Codex reads AGENTS.md with a 32 KiB ceiling against Claude's 4 MiB, and only the Claude provider rotates transcripts. Switching a group carries the memory/ tree across but not in-flight conversation context, and changes only take effect on the next container spawn.
Pointing NanoClaw at Kunavo
Kunavo publishes configuration references for this shape of client — the Claude Agent SDK and Claude Code guides cover the default runtime, and OpenCode and Codex CLI cover the two extended ones. Those are published setup references, not compatibility tests: NanoClaw has not been runtime-tested here, and everything above is a reading of its repository and documentation. Keep a working route available while you try it, run one bounded task, then read what your account recorded. Create a Kunavo account when you are ready to fund a key.
Choosing between agents rather than providers? NanoClaw vs OpenClaw compares isolation and provider freedom, and OpenCode pricing, Codex pricing and OpenRouter alternatives cover the three routes this page treats as add-ons. Claude Code pricing covers the default runtime's own billing in more detail.
FAQ
How much does NanoClaw cost?
NanoClaw itself is $0. Its homepage FAQ says it is free and open source under the MIT license and that your agent provider may charge for model usage, the nanocoai/nanoclaw repository carries an MIT license, and the documentation site's own page index lists 56 pages with no pricing, billing or plans page among them (checked September 21, 2026). There are no user accounts and no paid tier. What you actually budget is three things: the model API bill from whichever provider you attach, a Docker Desktop subscription if your company is above Docker's free-use thresholds, and the hardware if you run models locally. A different project, Clawland-AI/nanoclaw, is a Python agent for single-board computers and its hardware costs are not this product's.
Can NanoClaw use a custom API endpoint?
Yes, through one documented route: set NANOCLAW_ANTHROPIC_BASE_URL and NANOCLAW_ANTHROPIC_AUTH_TOKEN (or the matching --anthropic-base-url and --anthropic-auth-token flags) before running setup. Both are required together, the branch only runs when no Anthropic secret exists yet, and the base URL must be an origin with no /v1 path because the Anthropic SDKs append /v1/messages themselves — Kunavo's own ANTHROPIC_BASE_URL reference covers that trap. Setup then stores the token in the OneCLI vault keyed to that hostname, writes only the base URL into .env, and registers the custom-endpoint provider. The two-line .env snippet in NanoClaw's README is described there as being for one-off experiments, and on current main nothing reads an ANTHROPIC_AUTH_TOKEN value back out of .env to send upstream — the Claude provider asks .env for ANTHROPIC_BASE_URL only and hardcodes the token to placeholder — so that snippet is not the production path. Kunavo has not runtime-tested this route; it is a reading of NanoClaw's own documentation and source.
Does NanoClaw work with OpenRouter?
The route that NanoClaw's own documentation supports is /add-opencode with OpenRouter selected as the OpenCode backend, not an Anthropic base-URL override — OpenCode reads OPENCODE_PROVIDER, OPENCODE_MODEL and OPENCODE_BASE_URL, and its keys live in the OneCLI vault rather than .env. OpenRouter's own works-with page for NanoClaw (openrouter.ai/works-with-openrouter/nanoclaw, checked September 21, 2026) is a three-step directory placeholder — create an account, add credits and make a key, then configure it in NanoClaw — that names no environment variable, no config file and no model id, and it still links the old qwibitai address, so it is not evidence of a tested integration. On cost, OpenRouter's FAQ states it applies no markup on inference and charges 5.5% with a $0.80 minimum on card credit purchases, 5% on crypto, and 5% on bring-your-own-key usage above the monthly free allowance (checked September 21, 2026).
What is the best API for NanoClaw?
It depends on which runtime you are paying for, because NanoClaw's default is not a general provider list. Out of the box it runs Claude Code inside the container through the Anthropic Claude Agent SDK, so the choice is between an Anthropic API key, a Claude subscription token, or an Anthropic-compatible endpoint registered through setup. Beyond that, the providers guide says setup's runtime picker can install Codex inline, while OpenCode is added after setup with /add-opencode and the Ollama workflow is marked legacy — each alternative carrying its own credential surface, its own pinned CLI version and its own maintenance. A direct vendor key wins when you want the vendor's own caching and batch discounts. A gateway wins when you want one key and one balance across model families. A subscription wins on flat-rate heavy daily use. A local model wins for private or small work with no per-request charge, but NanoClaw's documentation currently marks its Ollama skill as legacy and not a supported drop-in switch on current main.
What is the cheapest API for NanoClaw?
Cheapest listed rate and cheapest way to finish the work are different questions, and for a long-running messaging agent the gap between them is mostly caching. Anthropic's pricing page puts a cache read at 0.1x the base input price and a five-minute cache write at 1.25x for the Claude models this page prices, so on an agent that re-sends a large stable prompt every turn, cache behaviour moves the bill more than a one-step change in model tier does. One documented route gives that up: NanoClaw's own Ollama notes (docs/ollama.md) say the Claude Agent SDK prepends a per-request hash that defeats Ollama's prompt cache unless you run a filtering proxy in front of it — and the same notes say Codex and OpenCode never emit that hash. Pick the least expensive model that finishes your turns without retries, then read what your provider account actually recorded.
What is the best model for NanoClaw?
On the default Claude provider a group's model accepts an alias — sonnet, opus or haiku — or a full model ID, with effort set to low, medium, high, xhigh or max; an unset model falls back to NANOCLAW_DEFAULT_MODEL and then the SDK default. That makes model choice a per-group setting you can tune per agent rather than one global decision. Two practical notes on the money. Kunavo's catalog lists Claude Sonnet 5 and Claude Opus 5 at the same per-token rate, $2.00 in and $10.00 out per million, so on that shelf moving between them is a capability decision rather than a price step. And NANOCLAW_FAST_MODE, documented as enabling Claude's fast serving tier for the values 1 or true, is explicitly noted as something that may change provider billing — Anthropic prices that tier, which its pricing page describes as being in research preview, at $10 per million input and $50 per million output for Claude Opus 5 and Claude Opus 4.8, on the first-party Claude API only (checked September 21, 2026). Both NANOCLAW_DEFAULT_MODEL and NANOCLAW_FAST_MODE sit under the changelog's Unreleased heading on main and are not part of the v2.3.0 release.
Fetched directly while writing this page, on September 21, 2026: the nanocoai/nanoclaw repository record and release list, package.json, container/cli-tools.json, the README, the hardened-image doc, nanoclaw.dev, the documentation index, the credentials, environment-variable and providers pages, docs/ollama.md, the changelog, the /add-opencode and /add-codex skill files, the OneCLI repository and Anthropic's pricing page. The Docker, Claude subscription, ChatGPT and OpenRouter amounts are quoted from those vendors' own pricing pages as recorded on the same date rather than re-fetched here, so confirm any of them in the vendor's checkout. Kunavo token rates come from the live catalog. Nothing on this page was runtime-tested against a NanoClaw install, and every dollar figure is illustrative token arithmetic.