VERROCCHIO

A WebMCP workshop that preserves missing proof, the next action, and human checkpoints for interrupted solo builders.

Comment

What VERROCCHIO does

VERROCCHIO is a WebMCP workshop for time-constrained solo builders whose projects are repeatedly interrupted. It keeps the next action, missing proof, active work, and human checkpoints together so an AI can help without turning an unverified claim into “done.”

The browser exposes four bounded, state-aware WebMCP tools:

  • inspect_workshop reads the live mission, missing proof (MANCA), active work, and the next bounded action.
  • call_fermo stops when evidence, scope, safety, or intent is ambiguous.
  • propose_workshop_draft validates and returns an unsigned plan that waits for human FIRMA.
  • return_work_result records a structured CLAIMED result without marking its evidence verified.

Human-only actions—FIRMA approval, evidence verification, resume, publication, and final submission—are deliberately absent from the WebMCP tool surface.

Why WebMCP is a strong fit

A person and an agent must act on the same live project state. With ordinary DOM automation, an agent has to rediscover controls and infer whether visible text is a claim, approval, or proof. WebMCP gives the page an explicit contract and exposes only the transitions the agent is allowed to perform at that moment.

The result is not simply fewer clicks. It is a clearer division of responsibility: the agent can inspect, draft, stop, and return bounded work; the human retains purpose, approval, evidence judgment, publication, and submission.

Better experience for interrupted solo work

A solo builder working around a full-time job carries coordination that a conventional team would normally distribute: rules, next actions, interruption recovery, evidence review, and stops before irreversible actions. VERROCCHIO makes that coordination visible and resumable.

The core flow is:

MANCA → unsigned plan → human FIRMA → bounded work → CLAIMED
      → human VERIFIED or CHANGES_REQUESTED

After an interruption, the builder can reopen the workshop and see what is still missing, what is active, and exactly where human attention is required.

What people and agents can now do together

The agent works directly with structured browser state instead of guessing from screenshots or page text. It can propose the smallest valid next plan, detect ambiguity, and return a result with checks, evidence references, and remaining risk. The human can adopt or reject that plan and verify the claim without surrendering the decisions that matter.

Implementation

VERROCCHIO is a React and Vite application with browser-local persistence. Tools are registered through document.modelContext.registerTool; their availability changes with workshop phase. Every mutation requires the state version returned by inspection and an idempotency key. Stale state is rejected, identical retries reuse the existing receipt, and conflicting reuse fails.

The current ChatGPT or Codex host supplies reasoning, so the app does not require its own OpenAI API key.

Evidence

The WebMCP-versus-DOM evaluation showed fewer actions for the tested stop and unsigned-plan paths (2 versus 3, and 2 versus 4). Recorded elapsed time was not faster, so no speed claim is made.

Existing-project disclosure

The July 19 baseline 91315a1 already contained the local workshop, MANCA proof gates, work packets, and human FIRMA/FERMO controls. Beginning with challenge-period commit 0176211 on August 27, the project added the four native WebMCP tools, state-aware availability, versioned and idempotent mutations, unsigned plans, CLAIMED results that cannot close proof, evaluation receipts, and the hosted judge path.

Comparison: https://github.com/honabochi/verrocchio-public/compare/91315a120059cae3b9eb68566c81d5c8d87fd903...webmcp-submission-2026-09-03

Known limitations

Persistence is browser-local rather than synchronized across devices. Human-only controls are a workflow boundary, not actor authentication. The keyless path depends on a WebMCP-capable host.

Built With

Share this project:

Updates

Submission history