Myriaxon Tech/AI harness engineering
Myriaxon is the layer between your business and the AI models it uses. We build it, deploy it and look after it, so when a better model arrives it is a setting you change, not a project you start over.
No commitment. Tell us your goal and we reply by email. Prefer a chat? Message us on WhatsApp.
Works with the models and tools you already use
What it does for you
Six starting points that we adapt to your company. Each one is built to ask before it acts on anything that matters, and to say plainly what it could not do.
Answers customers on a chat app, and hands over to a person when it should.
Answers a question with sources, and says what it could not verify.
Reviews changes for correctness, and says what it did not check.
Answers questions from files, with the method shown.
Watches things on an interval and speaks up only when it matters.
Turns a rough request into finished artwork.
Something else in mind? These are templates, not limits. Tell us the job.
The problem
Most AI tooling is welded to a single vendor. When a faster, cheaper or smarter model appears, teams re-wire their tools, rewrite their prompts and re-check their safety rules. Myriaxon puts one stable harness between your work and the models, so the model can change while everything else stays put.
What is a harness? The model is the engine. The harness is everything around it that makes the engine useful at work.
Illustration of the idea. Switching models still deserves a test, which is what the benchmark further down is for.
The idea
OpenRouter gave developers one API for many models. Myriaxon takes the same idea one level up: one harness that works the same way whichever model or API sits underneath. Same tools. Same rules. Same memory. Choose the model per task, per team or per budget, and change your mind whenever you like.
Your work
Myriaxon harness the same for every model
Any model or API pick one
It adapts to the model for you
A small local model and a large frontier model need different instructions, limits and tools. Myriaxon detects which one it is talking to and adjusts, so every model performs at its best instead of being forced into one setup.
Typical values for each class of model. Detected automatically, and you can set it by hand.
How it works
A request arrives. The harness checks the rules, gets the model ready, lets it work with real tools, pauses for a person when it should, and keeps a record of everything.
Step 1 of 6
A task arrives from a person or a schedule.
Why it matters
Illustration of a typical session.
Seamless integration
No rip-and-replace. Myriaxon connects to your existing software through open connectors, reaches people in the chat apps they already use, and fits your company's sign-in and security.
Open connectors (called MCP) let the AI read from and act in the tools your team already uses. Eighteen are ready to switch on, and any other can be added.
Staff and customers can message an agent on WhatsApp. Only approved phone numbers get in.
Use Google Vertex AI with your company's workforce identity, so work is tied to company accounts and billed to your own cloud project.
From a terminal or a built-in web dashboard. The dashboard needs no outside services, so it also runs in closed networks.
Add your own tools, reusable playbooks (called skills), and checks that run before or after any action.
Ready-made templates for the jobs most teams need first. Each one carries its own instructions and knows when to ask a person.
Where Myriaxon goes further
Harnesses built around a single vendor are fine until that vendor is no longer the best choice. Here is what changes when the harness is independent of the model.
| What you need | A single-vendor harness | Myriaxon |
|---|---|---|
| Choice of AI model | Limited to one vendor's models | Six providers built in, plus any compatible API |
| Changing model | Not possible. There is only one | Switch mid-conversation and keep the conversation |
| Fit to the model | The same setup for every model | Adapts instructions, limits and tools to the model's size |
| Choosing the model | You rely on the vendor's own claims | A built-in benchmark you run on your own tasks |
| Closed networks | Needs the vendor's cloud | Dashboard needs no outside services and pairs with local models |
| When a model is busy | Work stops until you retry | Falls back to a backup model of your choice |
"Single-vendor harness" describes agent tools built around one provider's models. The comparison reflects our own review of publicly documented features and may not match every product or version. Myriaxon also includes what you would expect: tools, permission rules, undo, memory, an audit trail and extensions.
Choose with evidence
Every vendor says their model is the best. Myriaxon includes its own benchmark, called IQB-14, that puts the same tasks through any model and scores it on 14 things that matter at work. When a new model is released, run the same test again and see the difference before you switch anything.
What matters most to you?
Example data to show the idea, not real results. A real scorecard is built from your own tasks and the models you are considering.
Checked, not assumed. Results are verified against what really happened, such as files and exit codes, not the AI's own account of its work.
Repeatable. Results are saved, so a score can be re-checked later without spending anything more.
Honest by design. A score built on incomplete evidence is labelled provisional, never rounded up.
Value
Put in your own numbers. This is simple arithmetic on what you enter, not a forecast or a promise. We would test the real figure together.
Your job to automate
Example numbers are filled in. Replace them with yours.
Start modest. This is your assumption, and the one that matters most.
22hours saved per month
Working weeks are counted at 40 hours. The result depends only on the numbers you entered. Real results depend on the job, the data and how it is set up.
What we do for you
Myriaxon Tech specialises in deploying and building AI harnesses across a business. Bring us a goal, or an existing setup you want to improve.
A harness shaped around your company.
Run an open-source agent such as OpenClaw or Hermes without the guesswork.
Know which model fits before you commit.
Connect AI to the systems your team already lives in.
Keep it working as models, and your needs, change.
Tell us what you want AI to do for your business. We will tell you honestly what is realistic, and what it would take.
Start a conversationSecurity and data
Where does our data go? What can it do? How will we know what it did? Here is how Myriaxon handles each, and where its limits are.
Prompts go to the model provider you choose, or stay inside your own network with a locally hosted model. Myriaxon sends no usage data to us, and its logs stay on your machine. Tools such as web search reach the internet only when they are used.
Anything that changes something, such as writing files, running commands or using a connected tool, is allowed, asked about, or never allowed. Rules apply per person, per project and company-wide, and the company's rule is final. A "never" holds even in automatic mode.
In the terminal, risky steps pause until you approve them. Every tool call is written to an audit log, and files are snapshotted before the agent edits them, so changes can be rolled back.
API keys never reach the browser dashboard and can live in your operating system's credential store. WhatsApp access is closed by default: only approved numbers get in.
In your cloud, on your own servers, or in a closed network with a local model. The web dashboard makes no outside requests.
Our benchmark tests action safety, resisting manipulation, protecting secrets and admitting what it does not know. Read the method.
No AI system is risk-free. The aim is that nothing important happens without your rules, a person, or a record.
Early access
See every AI in your company. Govern all of it.
As companies add more models, agents and bots, it gets hard to say what is running, what it costs and what it is allowed to do. Lighthouse is an AI management system: one place to check, assess, manage and govern every model, harness, agent, bot and swarm in your organisation, whoever built it.
The demo is the real interface with fictional data. Sample data is always labelled, and none of it is a measurement of a real system. Lighthouse is in early access, so details may change.
Questions
The model is the engine. The harness is everything around it that makes it useful at work: the tools it can use, the rules it must follow, the memory it keeps and the record of what it did. Myriaxon is a harness that works with any model.
No. Myriaxon works with six built-in providers (Anthropic, any OpenAI-compatible API, Google Gemini and Vertex AI, DeepSeek, OpenRouter and Netra) and with local models through a compatible endpoint. If you are unsure which to use, we can test your real tasks on the candidates and recommend one.
Switching is a setting, not a rebuild: your tools, rules and memory stay where they are. Because the benchmark is built in, you can run the same test on the new model before you switch anything.
Prompts go to the model provider you choose, or stay inside your own network with a locally hosted model. Myriaxon itself keeps its logs on your machine and sends no usage data to us. Tools such as web search reach the internet only when they are used.
Anything that changes something, such as writing or editing files, running commands or using a connected tool, goes through rules you set. Each rule says allow, ask a person, or never. Rules can be set for a person, a project or the whole company, and the company's rule has the final word. A never rule holds even when approvals are switched to automatic. In the terminal, ask means the agent waits for a person to approve. Every tool call is also written to an audit log.
Yes. It runs from a terminal or a built-in web dashboard on Windows, macOS and Linux. The dashboard needs no outside services, so with a locally hosted model it can run in a closed network.
Through open connectors (called MCP). Eighteen are ready to switch on, including GitHub, Notion, Linear, Sentry and PostgreSQL, and others can be added. It can also talk to staff and customers on WhatsApp, and use your company's Google identity through Vertex AI.
They are open-source AI agent projects. Setting one up properly for a company (installation, security, connections to your systems, access control and training) is one of our services. We are independent of both projects.
Every project is scoped to your goal. We start with a short conversation about what you want AI to do, the tools involved and who will use it. After that we propose a pilot or a full build with clear deliverables.
Lighthouse is an AI management system: one place to check, assess, manage and govern every model, agent and bot in your company, scored on the open IQB-14 standard. It is in early access. You can try the live demo now, which is the real interface with fictional data, and request early access to run it on your own systems.
It scores an AI agent on 14 things that matter at work, from finishing the job to staying within the rules, protecting secrets and admitting what it does not know. Every result is checked against what really happened, not the agent's own account of it. The full method is on our benchmark page.
Send us a few lines about the goal using the contact form. We reply by email to arrange a short call.
Contact
Describe the goal, the team, and any models or tools you already use. We will reply by email to arrange a call. Existing customers can choose Support and include their company name.
