PicoClaw vs OpenClaw: small-device deployment and API costs

PicoClaw and OpenClaw are both free, MIT-licensed, self-hosted AI agents, so the choice is not a price comparison: it is whether the host is a small single-board Linux device or an ordinary 64-bit computer. PicoClaw is a single Go binary from Sipeed with prebuilt Linux releases for RISC-V, MIPS, LoongArch, 32-bit ARM and s390x. OpenClaw is a Node.js gateway that needs Node 24.16+ or 26.1+, and whose only published hardware sizing — its Raspberry Pi guide — asks for 1 GB RAM, 1 core, 500 MB free disk and a 64-bit OS. Both bill you the same way afterwards: $0 for the software, and model API tokens for every turn.

Versions checked September 21, 2026. PicoClaw's latest stable release is v0.3.1, published July 3, 2026 (GitHub releases API) — not the v0.2.9 its README news section still ends on. OpenClaw's current release is 2026.9.5, published September 19, 2026, with extended-stable at 2026.7.35 (npm dist-tags); the GitHub list looks out of order because 2026.7.35 is a backport channel, not a newer build.

Confirm you are looking at the right two projects

This comparison has an unusually high name-collision rate, and the first three traps below hand you wrong numbers rather than no numbers.

• PicoClaw forks with identical descriptions. A GitHub search returns several repositories whose description string is byte-identical to the upstream one. Only github.com/sipeed/picoclaw is the project; a fork's main and release tags are its own. Check the owner reads sipeed before quoting a version.
• Look-alike domains and tokens. PicoClaw's README states that the only official website is picoclaw.io and the company site is sipeed.com, and that it has issued no tokens or cryptocurrency, calling trading-platform listings scams. When we checked at 14:38 UTC on September 21, 2026, picoclaw.io was serving a TLS certificate that expired September 10, 2026, so curl without -k failed verification. A certificate can be renewed at any moment, but the durable advice is to install from GitHub Releases rather than the one-click download.
• "Pico" the microcontroller. A Raspberry Pi Pico, RP2040 or RP2350 board does not run Linux, and PicoClaw's stated minimum is Linux kernel 3.x or newer — so the name does not imply that class of board.
• OpenClaw's earlier names. Clawdbot and Moltbot are the same product: OpenClaw's own lore page records Warelay, then Clawdbot, then Moltbot on January 27, 2026, then OpenClaw on January 30, 2026. Do not treat a "Moltbot vs PicoClaw" post as a third data point.
• The neighbours. NanoBot (PicoClaw's stated inspiration), NanoClaw, ZeroClaw and IronClaw are separate projects with their own prices and docs — see Nanobot vs OpenClaw and OpenClaw alternatives rather than substituting their figures here.

Who should pick which

The short version: pick PicoClaw when the device is the constraint and you can accept pre-1.0 software — its README says it "is in early rapid development" and to "not deploy to production before v1.0". Pick OpenClaw when the host is ordinary and the channel, plugin and companion-app surface is what you are buying. Neither choice changes the per-token rate you pay.

Hardware is the real difference, and the evidence for it is the build list

The strongest sourced difference is not a memory figure, it is which processors have a binary at all. The v0.3.1 release publishes Linux binaries for x86_64, arm64, armv6, armv7, riscv64, loong64, mipsle and s390x, alongside .deb and .rpm packages, Darwin, Windows, FreeBSD and NetBSD tarballs and an Android zip; its hardware doc separately lists verified x86, ARM, RISC-V, MIPS and LoongArch chips.

OpenClaw's position is narrower and stated in two separate places. Its own Node compatibility page says supported Node lines have no official Linux ARMv7 builds and directs you to a 64-bit OS. The RISC-V, MIPS and LoongArch case is a derivation rather than an OpenClaw statement: nodejs.org publishes no riscv64, mipsle or loong64 binary anywhere in its release index, and OpenClaw's docs never list those architectures — so there is no supported path, which is not the same as a demonstrated block. OpenClaw also gates the runtime further, requiring a node:sqlite binding that loads a WAL-safe SQLite library, and refuses a build inside the supported version range if that probe fails.

Read from PicoClaw's hardware compatibility doc, OpenClaw's install page and OpenClaw's Raspberry Pi guide on September 21, 2026. Three caveats belong next to these numbers rather than in a footnote. PicoClaw's own v0.3.1 tarballs are 16–24 MB compressed, which is larger than the 20MB storage minimum its docs state — the tarball may bundle the launcher and Web UI alongside the core binary, so size the card generously rather than assuming either number is wrong. PicoClaw footnotes its own memory headline, saying recent builds may use 10-20MB after a run of merged pull requests. And its configuration doc records a known limitation of the exec guard — it inspects only the command PicoClaw launches directly, not the child processes spawned by allowed tools such as make, go run or npm run — and advises that you "run PicoClaw inside a container or VM if you need stronger isolation than the built-in guard provides", which is precisely what a 64MB board cannot run.

Whose numbers are whose

PicoClaw's README carries a comparison table that assigns figures to OpenClaw. Those figures appear nowhere in OpenClaw's own documentation, so they are competitor marketing rather than a vendor spec, and this page labels them as such.

What it actually costs to run either one

The formula is the same for both: software $0 + (one-time board, PicoClaw path only) + tokens × rate + power or hosting. Neither project's documentation contains a per-token price table, so every rate below comes from the live Kunavo catalog rather than from either agent.

The figures below are illustrative token arithmetic, not measured task costs and not a bill ceiling. Assume a month of 3,000 agent turns, each sending 6,000 uncached input tokens and returning 500 output tokens — 18M input and 1.5M output in total. Your own turn shape will differ.

PicoClaw's documented cost lever is rule-based routing: its routing guide describes an agents.dispatch block that sends a turn to a configured light_model when a rule matches, with the first matching rule winning. Assuming an 80/20 split of the same month between Claude Haiku 4.5 and Claude Sonnet 4.6, the estimate is $14.28 against $30.60 for the whole month on Claude Sonnet 4.6. The 80/20 ratio is an assumption for illustration — neither project publishes a measured saving for its routing feature, and routing changes which entry serves a turn rather than the per-token rate.

The point that a small-board deployment does not change: PicoClaw's hardware doc lists network access as required, because every turn is a remote API call. A $10 board lowers the hardware line to roughly the price of the board; the token line above is untouched by it. 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. Cache charges and external tools sit outside this example, and the minimum top-up is $10 in prepaid credit, a funding minimum rather than a task fee or a subscription. See billing details.

Pointing either one at an OpenAI-compatible endpoint

Both snippets below are transcribed from the vendors' current documentation and were not executed against Kunavo's endpoint — Kunavo publishes no setup guide for either client and has runtime-tested neither. Treat them as a starting shape to verify with one small task, keeping a working route available while you do.

PicoClaw configures the whole thing inside one model_list entry, per its providers doc:

{
  "model_list": [
    {
      "model_name": "gateway-main",
      "provider": "openai",
      "model": "claude-sonnet-4-6",
      "api_base": "https://api.kunavo.com/v1",
      "api_keys": ["sk-kn-..."],
      "request_timeout": 300
    }
  ],
  "agents": { "defaults": { "model_name": "gateway-main" } }
}

OpenClaw needs the provider block and an explicit model entry, per its custom providers doc — its own wording is that agents.defaults.models "neither restricts overrides nor registers a new runtime model by itself":

{
  models: {
    providers: {
      kunavo: {
        baseUrl: "https://api.kunavo.com/v1",
        apiKey: "${KUNAVO_API_KEY}",
        api: "openai-completions",
        timeoutSeconds: 300,
        models: [
          {
            id: "claude-sonnet-4-6",
            name: "Claude Sonnet 4.6",
            input: ["text"],
            cost: { input: 1.2, output: 6, cacheRead: 0, cacheWrite: 0 }
          }
        ]
      }
    }
  },
  agents: { defaults: { model: { primary: "kunavo/claude-sonnet-4-6" } } }
}

The cost block above is worth typing. OpenClaw's custom-provider doc says an omitted cost block defaults to { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 }, and its usage-tracking doc says the Control UI Usage view shows a session-derived estimated cost — so an estimate built on zeroed metadata has no rate to price against. The zero defaults are documented; the resulting readout is a derivation from them, not a figure OpenClaw publishes. An omitted contextWindow stays unset, and context-budget callers fall back to 200,000 tokens only when neither discovery nor per-model metadata supplies one. Either way the authoritative number is your gateway's own usage ledger, not the agent's arithmetic.

Neither project gates any of this behind a paid plan, because neither sells one. For the wire-format background see OpenAI-compatible API, and the Anthropic base URL doc if you take the Messages route instead.

What moving between them actually costs you

Neither documentation set contains an importer in either direction — PicoClaw's only documented migration is picoclaw migrate for its own older versions, and OpenClaw documents no export to PicoClaw. That is absence of evidence rather than proof that no community converter exists, but plan on re-authoring by hand: the config formats do not translate (PicoClaw uses a flat JSON model_list plus a separate .security.yml for keys, OpenClaw uses JSON5 with models.providers.<id> and model aliases), and channel credentials, schedules and skills are re-created per platform.

Moving down to PicoClaw also costs capability: OpenClaw's iOS app, companion nodes, Signal, iMessage, Microsoft Teams, Google Chat and the plugin SDK have no PicoClaw counterpart in its docs. Moving up to OpenClaw costs the architectures — a RISC-V, MIPS, LoongArch or ARMv7 device has no supported OpenClaw path. PicoClaw declares a ClawHub registry and clawhub.ai markets skills for OpenClaw, but neither side publishes a compatibility statement and nothing was installed or executed for this page, so do not assume a skill carries across.

If the underlying question is which API to put behind whichever agent you pick, best API for OpenClaw and OpenClaw pricing cover the provider side, and creating a Kunavo account gets you one key and one prepaid balance to point either config block at.

FAQ

Checked September 21, 2026 against primary sources only: the GitHub API for both repositories and PicoClaw's releases, npm dist-tags for OpenClaw, both projects' READMEs, PicoClaw's hardware, providers, routing, configuration and security docs, and OpenClaw's install, Node compatibility, Raspberry Pi, channels, pairing, security, sandboxing, agent-runtimes, usage-tracking and model-provider pages. The picoclaw.io certificate was observed expired at 14:38 UTC that day. Neither client was runtime-tested against Kunavo, no benchmark was reproduced, and hardware retail prices were not independently verified. Kunavo token rates read from the live catalog; every dollar figure here is illustrative token arithmetic.