Days 3–4 of building WebMCP-QCG
I slept before Day 3.
That sounds like a small achievement in a development journal, but it changed the architecture. My most useful optimization was giving the human operator a full night of RAM.
I returned to WebMCP-QCG rested enough to see the project clearly. I had begun with an ambitious quantum router spanning frameworks, simulators, providers, notebooks, processors, and several futures at once. The map was exciting. The product was hiding inside it.
Then one question gave the project its purpose:
Before an agent asks a quantum system to do something, has that call earned the right to run?
That question turned the router into a gate.
One artifact enters the browser
The transaction begins with a real Q# file selected by the human.
QCG reads it locally, computes its SHA-256 digest, records its provenance, and compiles it through a bounded browser Worker. The agent receives a structured artifact identity rather than a loose description of the experiment. It can inspect the manifest and evaluate the intended action against explicit limits.
The result is one recommendation:
- reuse fresh evidence;
- reject an unsupported call;
- recompile for the selected target;
- simulate locally first; or
- mark the experiment technically ready for a later external workflow.
Technical readiness is a statement about evidence. Authority is a decision made by a person.
The moment the machine waited
My favorite state is simulate_first.
The recommendation appears. The reason codes are visible. The proposed limits are visible. Then the system waits.
That pause is the product.
If I accept, QCG creates a short-lived, single-use consent for one bounded local simulation. The Q# Worker runs the experiment, returns its histogram, and consumes the consent. If I defer, the experiment remains where it is. If I override the recommendation, the receipt preserves my choice and justification instead of rewriting the history into a cleaner story.
In the validated Bell path, the browser completed 64 correlated shots. The effect counters recorded one local simulation and zero provider or QPU submissions. The receipt linked the artifact digest, compiler version, policy, recommendation, human decision, and measured effect.
For the first time, I could follow the whole chain and explain every transition.
Four tools with four responsibilities
WebMCP-QCG now exposes a deliberately small surface:
inspect_quantum_experimentevaluate_quantum_callrun_bounded_qsharp_simulationexport_quantum_evidence_report
The third tool only appears when the recommendation, application state, and visible human consent all agree. The fourth exports what already happened; it does not silently re-evaluate the experiment.
This separation gives each tool one responsibility and gives the browser a history that both a human and an agent can inspect.
The current codebase passes 34 automated tests across six files and builds 127 TypeScript/Vite modules. Those numbers matter because they describe the proof I actually have. They also reveal the next work: live browser lifecycle testing, the optional QCG panel in Chrome DevTools, deployment parity, and a clean multi-agent handoff.
A room for agents, with one human gate
I am now extending this discipline into DevTools.
The goal is a shared debugging room where Codex can map contracts and regression risks, a Gemini-oriented client can challenge the browser diagnosis, and I can compare both contributions. Each observation will keep its declared actor, evidence references, confidence, and review state.
The collaboration tools may read context, post an observation, request human review, and export a handoff. Quantum consent and execution remain in the primary QCG workflow.
This is the future of browser agents that interests me: several capable systems contributing evidence while their responsibilities stay legible. Intelligence can be distributed. Authority still needs an address.
Why Winter matters
The first article belonged to Autumn. It captured discovery: gold light, a closed gate, and ideas arriving faster than I could classify them.
Winter represents structure. Bare branches reveal the load-bearing parts. Footprints remain visible. Every feature must show the evidence that earned its place.
I am building this for more than a hackathon. I see a future where students, independent researchers, and developers with ordinary computers can prepare meaningful quantum work through a browser. Local simulation, cloud infrastructure, and physical processors will keep different roles. A clear gate can help people understand which path fits their experiment before cost, complexity, or automation carries them forward.
That matters to SecuredMe. It matters to the educational tools I want my children and future students to inherit. Powerful systems should teach judgment alongside capability.
At the end of Days 3 and 4, the project felt alive for a new reason. It had learned how to wait.
The browser can inspect. The policy can recommend. I decide when the gate opens.
