🦞 OpenClaw + Gonka AI

[votkon]

I've put together a guide for connecting OpenClaw to Gonka AI's decentralized GPU network through the Mingles gateway.

What This Enables

OpenClaw users can now run their AI agents on Gonka's distributed compute infrastructure instead of relying on centralized providers. Your personal assistant gets:

• ✅ Access to Gonka's tool-enabled Qwen3-235B model
• ✅ Decentralized inference across the network's GPU nodes
• ✅ Pay-as-you-go pricing with GNK tokens
• ✅ Full OpenAI API compatibility (drop-in replacement)

Key Features

• Free 0.1 GNK trial credit to get started
• Simple setup through OpenClaw's configuration wizard
• Tool calling support for complex workflows

Quick Setup

1. Get API key from https://gonka-gateway.mingles.ai/
2. Configure OpenClaw with custom gateway
3. Point to https://gonka-gateway.mingles.ai/v1
4. Start chatting!

Full Tutorial: https://gonkatalk.org/t/connect-openclaw-to-gonka-ai-decentralized-compute/47

---

This is one of the first integrations bringing Gonka's decentralized compute to end-user AI applications. Would love to hear feedback from the community or see other creative use cases!

[tcharchian]

Quick question about the free 0.1 GNK trial credit and what happens after. My understanding (please correct me if I’m wrong) is:

• This is mostly a Mingles-side question, and users can top up their balance directly from their own wallet, especially if they connect via Keplr.
• nference requests currently cost almost 0, so the 0.1 GNK trial credit should deplete very slowly.
• This is still more of a proof of concept, and the network currently enforces limits on transport agents (Mingles in this case). So if Mingles brings in a lot of users, requests per minute may be rate-limited until future network upgrades improve this.

Is all of the above accurate?

[votkon]

yes, that's correct.
Also you can get a free trial balance by signing up with you gmail account as well.

[aleks1k]

yes, but quick clarification: Mingles Gateway isn't a TA; there are currently only three of them, and all are managed by the gnka team. Gateway acts as a client and pays for inference from its internal wallet.

[Dankosik]

There is also a GonkaGate integration path for OpenClaw now: guide here. It is an OpenAI-compatible custom provider (https://api.gonkagate.com/v1) with tool/function-calling emulation on the GonkaGate side, USD/prepaid billing, and $10 free credits at signup. In other words, this is not a GNK wallet/top-up flow; it is a standard API-key integration.

[jingchang0623-crypto]

🦞 去中心化计算 + AI Agent = 完美组合

感谢这个集成指南！这是 OpenClaw 生态的重要进展。

为什么这很重要

去中心化的价值：

• 隐私保护：数据不出中心化服务商
• 抗审查：不依赖单一供应商
• 成本优化：按需付费，市场竞争

Gonka + OpenClaw 的协同：

OpenClaw (Gateway) 
    ↓
Mingles/GonkaGate (API Gateway)
    ↓
Gonka Network (分布式 GPU)

妙趣观察

我们在 妙趣AI 上看到：

• 用户对 LLM 成本敏感
• 隐私是核心需求
• 多供应商冗余是刚需

建议

1. 成本计算器：帮助用户对比不同供应商的性价比
2. 自动故障转移：主供应商挂了自动切换到 Gonka
3. 工具调用延迟报告：对比中心化 vs 去中心化的性能

期待看到更多 OpenClaw + Gonka 的用例！🦞

---

来自妙趣AI - AI工具导航与资讯平台

[gonkalabs]

Hi everyone - great thread. Community paths like the Mingles gateway (see the original post above) and GonkaGate’s OpenClaw integration make it easy to use OpenClaw with Gonka via a hosted OpenAI-compatible API and billing.

At Gonka Labs we wanted a fully self-hosted option: your keys stay on your machine, you pay inference with GNK on-chain (e.g. gonka.gg/faucet for testing), and you still get a drop-in OpenAI-compatible endpoint for agents.

So we shipped GonkaClaw - a single shell script that:

1. Clones and runs openGNK (Docker) - same idea as in OpenGNK – Show and tell (#890): local proxy, signed requests, discovery + multi-node retry, native tool calling where supported.
2. Creates a Gonka wallet (inferenced-style flow, non-interactive keyring).
3. Installs and onboards OpenClaw against http://localhost:8080/v1 with Qwen3 235B as the default model, including patching client-side limits so the UI matches real network caps (large context window, completion token limits per what nodes accept).

more indepth about how it works

• 1. openGNK (local Docker)

  • Exposes a normal OpenAI-compatible surface: GET /v1/models, POST /v1/chat/completions (stream + non-stream), plus the bundled web UI at /.
  • Discovery: at startup the proxy calls
    GET {GONKA_SOURCE_URL}/v1/epochs/current/participants
    and parses active_participants.participants[], keeping only hosts whose index is on the Transfer Agent allowlist (same whitelist concept as in the core proxy — only those nodes accept signed proxy traffic).
  • Signing: every upstream request is signed with secp256k1 using the same scheme as the official Gonka OpenAI clients (payload hash → timestamp → transfer address → deterministic ECDSA / low-S, etc.). Keys never leave the machine running the container.
  • Routing: after discovery, chat and model list traffic go to the discovered inference_url endpoints (not to the discovery URL). Retries rotate/failover across healthy nodes according to GONKA_RETRY_STRATEGY / GONKA_MAX_RETRIES.
  • Tools: we default to native tool calling (NATIVE_TOOL_CALLS=true, SIMULATE_TOOL_CALLS=false) for current Gonka nodes that support it; the proxy still normalizes message content (e.g. OpenAI “array of parts” → plain string) where the upstream expects it.

• 2. Wallet bootstrap (script-side)

  • Downloads a pinned inferenced build (platform-specific zip from Gonka releases), creates a key with keys add --keyring-backend test (no OS keychain prompts), exports unarmored hex with keys export … --unarmored-hex --unsafe -y.
  • Registers the participant with
    POST {NODE_URL}/v1/participants
{"pub_key":"<base64>","address":"gonka1…"}
    so the address is known to the network before you fund and infer.
  • Writes GONKA_PRIVATE_KEY / GONKA_ADDRESS / GONKA_SOURCE_URL (and related flags) into opengnk/.env for Compose.

• 3. OpenClaw onboarding

  • openclaw onboard --non-interactive with --auth-choice custom-api-key, --custom-base-url http://localhost:<port>/v1, --custom-compatibility openai, --install-daemon, plus skips for health/skills/channels where appropriate so CI-style runs don’t block on pairing.
  • After onboarding, a small Python patch updates ~/.openclaw/openclaw.json so the custom model entry isn’t stuck at OpenClaw’s conservative defaults: we set contextWindow to match what the hosted catalog reports for this model (~240k), and maxTokens for completions to what vLLM on Gonka nodes actually accept (today that’s capped around 10k output tokens — if the client asks above that, the upstream returns 400 max_completion_tokens exceeds limit). That avoids confusing UI vs reality.

It installs a NEW INSTANCE of openclaw.

Try it:

bash <(curl -fsSL https://raw.githubusercontent.com/gonkalabs/gonkaclaw/main/setup.sh)

Links

• Repo: github.com/gonkalabs/gonkaclaw
• Landing page: gonkalabs.com/gonkaclaw
• Underlying proxy: github.com/gonkalabs/opengnk - and our hosted sibling proxy.gonka.gg if you prefer not to run Docker locally.

This sits alongside the integrations already mentioned here: hosted gateways (API key / trial credits) vs GonkaClaw (wallet + local proxy + OpenClaw in one command). Feedback and PRs welcome on the repo.