Skip to content
MCP / BW-001 · CONNECTED

Durable intent, ownership, and proof

Give your coding agent
a real contract.

Benchwright turns product intent into durable features, behavioral claims, responsible systems, and executable tests. It connects through MCP to the coding agent you already use, so every change starts with the right context and ends with proof.

Walk through the workflow Request early access
You bringYour coding agent
You doSet intent + approve
Benchwright keepsContext + evidence
Your coding workflowConnected through MCP
Your workflowYour coding agent

Describe, review, and build in the tool you already use.

Durable product recordBenchwright

Approved intent, ownership, and proof remain durable.

How you interact

  1. 01Describe

    State the outcome.

  2. 02Review

    Challenge the structure.

  3. 03Decide

    Resolve open choices.

  4. 04Approve

    Authorize the work.

Integrates directly with your harness and persists in the cloud.

A real workflow, inside your coding agent

Watch a product brief become build-ready context.

Step through a real CAD search example. See how your agent proposes structure, surfaces a product decision, records your approval, and carries exact claims and tests into implementation.

CAD search exampleStage 1 of 4: Describe

Brief → proposed contract

Describe the outcome in the agent you already use.

Benchwright is not another chat. Tell your coding agent what you want to build. Through MCP, it turns your brief into product intent, proposed structure, and explicit open decisions.

01
Your coding agentConversation
You
I'm building a visual CAD search tool: a user photographs an object and finds matching models. We plan to test on the ABC Dataset, use DINO for embeddings, and YOLO for object crops.
Your coding agent
Benchwright is connected. I'll separate durable product intent from implementation choices, propose the structure, and return only the decisions that need your judgment.

You never leave this conversation.

02
Durable contractSaved in Benchwright
Desired outcome
Find a matching CAD model from one or more photographs.

The agent extracts the durable outcome from your prompt.

Known ingredients
ABC Dataset · DINO · YOLO

Named technology is preserved without being mistaken for a complete product contract.

Needs a decision
Object choice · system ownership · evidence · definition of done

The agent can now ask what the technical brief leaves unresolved.

You stay in your coding agent. Benchwright becomes the durable record behind the conversation.

Shown inside your coding-agent conversation. Benchwright is the durable context layer connected through MCP; example records are simplified for clarity.

One workflow. Clear responsibilities.

You set the intent. Your agent does the work. Benchwright keeps both aligned.

01 / YOUR WORKFLOW

Stay in the conversation you already use.

Describe the outcome to your coding agent. It proposes structure and surfaces open decisions; you review and approve them. Benchwright preserves the result as the contract for implementation.

One conversationShared product context

The contract beneath the conversation

One model connects intent to code.

Benchwright records why the product exists, what users should observe, where each behavior is owned, and which tests prove it. Your agent retrieves only the slice relevant to the change.

01

Technical truth

Behavior, responsibility, and proof

Claim

A falsifiable statement about observable product behavior.

is owned by a System and verified by Tests

System

The architectural boundary responsible for keeping a Claim true.

bounds the code and context an agent must inspect

Test

Executable evidence that a Claim still holds.

maps to a named test function in code
02

Product truth

Intent and user-facing capability

Product

The durable outcome your organization intends to deliver.

breaks down into user-facing Features

Feature

A user-facing capability that persists beyond a ticket.

is defined by Claims that people can approve

One decision, carried into code

You make the call once. Benchwright keeps it true.

Your agent turns a short answer into observable behavior, responsible systems, and executable tests. Benchwright keeps that approved decision available through implementation and future changes.

Human decision / approved
“User selects the crop.”
01 / SYSTEMCAD Search Web
Feature
Multi-Photo Object Search
Claim
Display every YOLO candidate without selecting one automatically.
Test
test_web_displays_candidates_without_auto_selection
Passed · browser evidence
02 / SYSTEMCAD Search API
Feature
Multi-Photo Object Search
Claim
Embed only the crop explicitly selected by the user.
Test
test_search_embeds_only_selected_crops
Passed · API evidence
Committed product evidence43 linked Test IDs · 39 Python + 8 browser tests · 7,168 searchable models

Ready when your agent is

Build from decisions that survive the conversation.

Connect Benchwright through MCP. Keep your existing coding workflow, add durable product context, and give every change a clear definition of done.

Request early access