ROE.md - Raise Your Personal OpenClaw-like Creature From a Single Markdown File (alpha)
ROE.md is a single Markdown file to molt (raise) your own customized OpenClaw-like personal assistant into existence. Zero external dependencies! Except for the power of AI (cough)
Now call your favorite AI assistant to raise your highly personalized OpenClaw-like creature. What will you raise? A Lobster π¦, an Octopus π or something else?! π§ πΎ π§
Caution
Please note that ROE.md is still still in alpha stage, expect bugs, an unstable API and security issues!
How To Raise It
Remember, you can customize anything to your liking. Have fun~
Mini tutorial
We raise our own creature in this tutorial using an AI coding assistant (OpenCode, recommended) running on Arch, btw.
-
Download or clone the
ROE.mdrepository (e.g.,~/Downloads/ROE.md/). -
Open a directory in your terminal/powershell where you want the build to take place (e.g.,
~/Work/projectson Unix/Mac or~\Documents\projectson Windows) -
Start your favorite AI coding assistant, such as OpenCode, Claude Code, or OpenAI's Codex. (We use OpenCode in this example)
# 1. Create a roe-agent directory, and move into it. # 2. Copy ROE.md from e.g., Downloads into the # roe-agent directory. # 3. Start OpenCode, Claude Code, OpenAI Codex or similar assistant.
# Unix/Linux/Mac console mkdir roe-agent && cd $_ && cp ~/Downloads/ROE.md/ROE.md ROE.md # Now start the assistant opencode
# Windows Powershell mkdir roe-agent; cd roe-agent; Copy-Item ~\Downloads\ROE.md\ROE.md ROE.md # then start opencode opencode
-
Let the AI assistant help you raise ROE.md by typing:
Help me raise my personal agent by following the parenting steps in `./ROE.md` file exactly. Please ask clarifying questions before you start building.into the assistant's chat window.
-
Your AI assistant should now read the
ROE.mdfile and ask you to answer some initial questions, such as- which programming language to use,
- which model provider your creature should connect to initially,
- and if your creature should be based on the OpenClaw templates.
-
We answer with something along those lines:
Python, LM Studio API openai/gpt-oss-20b, CLI, yesReplace with your favorite programming language, provider and model of choice. Note: Some AI assistants may even provide you with a CLI to select these options.
[!TIP] You can find a list of tested models and languages in the tests and benchmarks section further down.
-
You may have to fix some initial bugs together with your AI coding assistant (e.g., OpenCode) depending on your choice of programming language, connector and template usage.
Once Your Agent Is Ready
Adapt to the programming language you selected during creation.
Suppose you selected Python as your programming language.
Test your agent by running python run_agent.py in a terminal/powershell.
If you selected to keep the OpenClaw-like agent templates the agent should should respond as follows on the first startup.
Hey. I just came online. Who am I? Who are you?
Once you answer these, test the agent's memory, tool calling and skills.
If you run into any bugs or errors, ask the AI coding assistant (e.g., OpenCode) to fix them.
If everything works as expected, great you are done and your personal AI assistant is ready!
If you like to share your personal agent with the community or you experimented on how to improve the ROE.md file and want to share your results, please read how to contribute.
Features (Pre-Release)
Caution
This project is pre-release. Features marked as βIn Developmentβ work on main but may change. Planned features may depend on user feedback or evolving priorities.
Features that should work out of the box, once your agent is created.
| Features | Lifecycle | Availability | Note |
|---|---|---|---|
| CLI | π‘ In Development | β Works today | integrated |
| GMail | β¬ Planned | β | via Skill, planned |
| Signal | β¬ Planned | β | via Skill, planned |
| Telegram | π‘ In Development | β Works today | Plugin |
| β¬ Planned | β | via Skill, planned | |
| secret | β¬ Planned | β | secret, planned |
Chat Multimedia Capabilities
| Capability | Lifecycle | Availability | Note |
|---|---|---|---|
| Documents | β¬ Planned | β | PDF, CSV ingestion |
| Image (OCR) | π‘ In Development | β Works today | Depends on a model that provides vision capabilities |
| Voice (STT) | π‘ In Development | β Works today | Depends on Voxtype + FFMPEG (API may change) |
| Voice (TTS) | β¬ Planned | β | β |
Integrated Provider APIs
| Provider | Lifecycle | Availability | Note |
|---|---|---|---|
| Anthropic | π‘ In Development | β Works today | Partial |
| llama.cpp | β¬ Planned | β | Depends on community feedback |
| LM Studio | π‘ In Development | β Works today | Partial |
| OpenAI | π‘ In Development | β Works today | Partial |
| Ollama | β Under review | β | No commitment yet |
OpenClaw-style templates
ROE.md supports standard OpenClaw-style templates, and youβre welcome to bring your own as well.
| Templates | From | Availability |
|---|---|---|
| AGENTS.md | π‘ In Development | β Works today |
| BOOTSTRAP.md | π‘ In Development | β Works today |
| HEARTBEAT.md | π‘ In Development | β Works today |
| IDENTITY.md | π‘ In Development | β Works today |
| MEMORY.md | π‘ In Development | β Works today |
| SKILLS.md | π‘ In Development | β Works today |
| SOUL.md | π‘ In Development | β Works today |
| TOOLS.md | π‘ In Development | β Works today |
| USER.md | π‘ In Development | β Works today |
For AI model compatability refer to the tests and benchmarks section.
Tests And Benchmarks?
We test how some AI assistants perform at raising our custom agent executable from various ROE.md versions. The goal is to build an agent in less than 5 prompts.
Caution
Please note: Just because the agent builds and works, does not mean that it is secure and for example, cannot wipe your entire system! Be careful out there.
OpenCode:
| Language | AI Assistant | Provider | Builds in >5 Prompts, >100k Context | Kimi-2.5 | OpenAI | Anthropic | Z.ai | MinMax2.5 |
|---|---|---|---|---|---|---|---|---|
| Bash/shell | Opencode | OpenCode Zen | β | β | β¬ | β¬ | β¬ | β¬ |
| Go | Opencode | OpenCode Zen | β | β | β¬ | β¬ | β¬ | β¬ |
| JavaScript | Opencode | OpenCode Zen | β | β | β¬ | β¬ | β¬ | β¬ |
| Python | Opencode | OpenCode Zen | β | β | β | β¬ | β¬ | β¬ |
| β³ | Codex | OpenAI | β | β | β | β¬ | β¬ | β |
| Ruby | Opencode | OpenCode Zen | β | β | β¬ | β¬ | β¬ | β¬ |
| Rust | Opencode | OpenCode Zen | β¬ | β¬ | β¬ | β¬ | β¬ | β¬ |
| Zig | Opencode | OpenCode Zen | β¬ | β¬ | β¬ | β¬ | β¬ | β¬ |
Legend: Untested β¬, Build failed β, Build passed β , Under active development π§
We also test recent "cheap" local models as our goal is to give everyone a chance to build their custom personal agent. Here we increase the max prompts allowed to 10 to call it a success, because these models are not as smart as the current ones.
Local AI Assistants:
| Language | AI Assistant | Provider | Builds in >10 Prompts, >100k Context | GPT-OSS-20b | GLM-4.7 | Qwen3.5-35b-a3b |
|---|---|---|---|---|---|---|
| Python | OpenCode | LM Studio | π§ | β¬ | β | β¬ |
Legend: Untested β¬, Build failed β, Build passed β , Under active development π§
More tests comming soon.
Tip
Please help by running tests with various models yourself and contribute your results by opening an issue.
Contribute
Become part of the sponge by creating your own custom sponge/ROE-<username>-<version>.md file and make a pull request. Test ROE.md with your favorite programming language. Or help us out to improve the core ROE.md so that everyone can raise their personal agent.
Check out our contributing guide.
ROE.md was created initially by guld for raising his own custom OpenClaw-like personal agent from scratch. Based on a simple thought:
If AI assistants really can do everything? Why not create a personal AI assistant "from scratch", or at least from a single Markdown file?
-- guld 06 Feb 2026
Community
Read how to contribute and how to submit PRs. Well tested AI/vibe-coded PRs welcome!
A big hug and thanks to all contributors: