Hailo-Ollama + OpenClaw Adapter for Raspberry Pi (AI HAT+2)

[sergiit]

Hello Hailo Community!

Thank you for the interest in this project! Please note that this was developed as a short-term Proof of Concept (PoC) to explore Hailo’s connection to OpenClaw.

Due to the rapid evolution of the ecosystem and shift toward focusing on Hailo’s core strengths—specifically Computer Vision and Whisper via MCP—this project is no longer being actively maintained. We recommend users “bring their own LLM” for OpenClaw’s agentic requirements.

The code is available for experimental use, and community forks are encouraged.

I want to share a small but useful bridge I built to get OpenClaw talking to Hailo-Ollama server (Hailo Model Zoo Gen AI) on Raspberry Pi 5 with Hailo accelerators (AI-HAT+2 in my case).

If your Raspberry Pi 5 is currently running Ollama and your CPU cooler is revving like it’s launching into orbit — you will need next setup!

The only thing standing in your way? Hailo-Ollama doesn’t speak the exact OpenAI /v1/chat/completions dialect OpenClaw expects by default.
This lightweight proxy https://github.com/tishyk/hailo-ollama-openclaw-adapter bridges the gap: it redirects requests, trims context to prevent overload, reformats responses, and makes everything work smoothly.

It’s very simple: redirects requests, trims context to last user message (to avoid overload on smaller models), reformats responses, and lets you chat via OpenClaw dashboard or Telegram/WhatsApp/etc.

Tested on:

• Raspberry Pi 5 (64-bit Raspberry Pi OS) with 8Gb of RAM

• Hailo AI HAT+2 / Hailo-10

• OpenClaw: 2026.2.9 (Manual is not working already for April releases)

• Hailo Model Zoo GenAI 5.2.0

• Python 3.13 in venv

Prerequisites (already assumed done)

• Hailo drivers + platform tools installed

• Hailo-Ollama server running on default port 8000 with at least one compatible model pulled (e.g. qwen:1.5b, smaller is better for speed)

• OpenClaw installed and dashboard accessible

If you’re missing any of those → check official Hailo docs or community threads for installation.

Updated setup guide for OpenClaw 2026.04.20 and Hailo MZ GenAI v5.3.0

Setup Steps for OpenClaw 2026.02.09 and Hailo MZ GenAI v5.2.0

1. Clone the repository

git clone https://github.com/tishyk/hailo-ollama-openclaw-adapter.git
cd hailo-ollama-openclaw-adapter

2. Create & activate virtual environment

python3 -m venv venv
source venv/bin/activate

3. Install dependencies

pip install -r requirements.txt

4. (Optional) Review / tweak adapter.py

Open adapter.py and check:

• hailo_url = “http://localhost:8000/api/chat” → change port if your Hailo-Ollama is not on 8000

• Context trimming (keeps only latest user messages)

• Response reformatting block

Defaults work for most cases.

5. Update your openclaw.json configuration file

Locate openclaw.json (usually ~/.openclaw/ or in install folder).

Add/update a provider like this (check openclaw_example.json).

…

Important: Use port 11435 (adapter listens there)

6. Restart OpenClaw gateway

openclaw gateway restart

7. Start the adapter (inside venv)

. venv/bin/activate
uvicorn adapter:app --host 0.0.0.0 --port 11435 --timeout-keep-alive 240 --limit-concurrency 2

8. Make sure Hailo-Ollama is running in a seperate terminal window

hailo-ollama

Confirm listening on http://localhost:8000

9. Test in OpenClaw Dashboard

• Open dashboard with command:

      openclaw dashboard

• Go to Agents / Models / Providers

• Look for hailo-ollama-local → should show as available

• Go to Chat

• Send test: Hello! Tell me a short joke about AI hats.

You should get fast, Hailo-accelerated response. Don’t forget - it will take longer for the first time hailo-ollama server started because of the model loading.

Troubleshooting

• Provider not visible → check JSON syntax + restart gateway

• Adapter gets no requests → wrong baseUrl (must be 127.0.0.1:11435)

• Timeouts from Hailo → try smaller model, check Hailo-Ollama logs

• Context too long → adapter already limits to last 2 messages

Final Notes

This is a minimal adapter — no streaming yet, no full tool calling support, but it gets basic chat working very nicely with Hailo acceleration and zero cloud. I tested it with qwen:1.5b model and telegram bot and found possible to put images or send voice over the Telegram bot to get answers.

Feel free to fork, PR improvements (streaming would be cool!), or ask questions here.

I’ll add screenshots and a short demo video later.

Thanks for reading!

[user704]

Good job！very fast in chat.

[user1031]

does anybody succeed to enable OpenClaw with full tools calling?

[Badz]

Hi

Thanks for this

Do you know what the context window of the various model ? I fill it at first request.

Thanks

[Pav1985]

Have you tried with model zoo 5.3.0? Looks like after last hailo update that doesn’t work anymore

[sergiit]

I see compatibility issues with latest openclaw version. Here are new openclaw.json setting needed with a profile changes to another file. I will try to update this guide soon.

[sergiit]

Sorry, I don’t know what is that. It was a weekend project

[sergiit]

Updated adapter installation flow, OpenClaw setup and user guide for 5.3.0 and OpenClaw 2026.04.20 version on my github GitHub - tishyk/hailo-ollama-openclaw-adapter: Detailed steps will be added later · GitHub

[Badz]

Thanks @sergiit . I already upgraded to openclaw 2026..4.23. is it even worth trying or it will break ? ( In the doc you say that concurrency changed)

Thanks

[sergiit]

Hi, you can try the latest version of adapter with steps described on the GitHub, it should work the same way as for older April releases. As I remember latest Openclaw had the same olama setup.

[Badz]

thanks @sergiit . I found a problem with \\` and found a workaround : i added an issue to your repo.

Can you make it use any tools ? I cant even ask it to read files on the filesystem: using quen3:1.7 , it keeps answering “I’m sorry, but I can’t provide the requested information.” unless I ask for the capital of france (which it correctly answered 8)

[Nico_Broeckers]

I reply to your New Version that you published from Github…

The adapter is good, but I’m running into a problem. The AI ​​agent is unable to access the PC to open programs there. Are you able to do this? Because, judging by your screenshot on from the newest Version on GitHub, it looks like it can.

I am using the Qwen3:1.7b model.

Compared to your previous version, the agent is able to retain more information. In your screenshot on Github, one can see that it answers what your operating system, that is Linux. This implies that it must be able to access your Raspberry Pi 5. I followed your tutorial instructions to the letter—including, of course, performing the update to version 5.3.0; otherwise, I wouldn’t be able to use the latest model.

One more question: Are you able to access the internet using the agents?

[Michael]

Hi all,

We are really happy with the interest of the community in Hailo solution for OpenClaw ecosystem.
Please stay tune - soon we will have exciting updates!

Thanks,

[sergiit]

Hi, I didn’t test tools, internet access and other OpenClaw functionality. There were a lot of issues I saw with a setup for OpenClaw latest versions, unfortunatelly no time to fix or test it. I think, there is a way for a better Hailo 10 chip usage with OpenClaw like with MCP Server or some tool.