2026-10-11 16:38 UTC

Quiet Grid Labs claims Viaduct exposes version-pinned architecture change sets through MCP with constraints, acceptance criteria, and commit-linked completion reports, making coding-agent work reviewable against an explicit system model.

state: seedheat: lowuncertainty: highconvergesscott: lowcoding-agents architecture-governance change-management mcpQuiet Grid LabsIgor Golovko

What is this?

Viaduct is a C4 architecture-modelling tool from Quiet Grid Labs, which describes itself as an engineering lab building focused developer tools. The supplied case describes a Show HN announcement and claims MCP-accessible, version-pinned architecture change sets with constraints, acceptance criteria, and commit-linked completion reports for coding agents. The relevant web snippet confirms the product and maker but does not establish those workflow details, Igor Golovko’s role, or whether the claimed reviewability has been demonstrated; the time-series AI product at viaduct.ai is a separate search result, not corroboration.

Why it matters to Scott

Viaduct’s claimed version-pinned task constraints and commit-linked reports converge with Scott’s Spec-Driven Development, Governance Trace, and Attributable named-agent runs, but currently offer another product example rather than a demonstrated extension or reason to change his builds. The supplied grounding does not verify those workflow details or independent conformance checking; the radar already follows adjacent architecture-review tooling on workbraid-visual-agent-proposals, not Viaduct itself.
ip:concept.spec-driven-developmentip:concept.governance-tracedev:concept.attributable-named-agent-runsradar:workbraid-visual-agent-proposalsradar:naeos-coding-agent-engineering-systemradar:coding-agent-self-report-failure-blindness
queries asked of Scott's wikis
  • coding agent architecture constraints and system-model drift
  • version-pinned specifications and agent task contracts
  • acceptance criteria commit provenance agent completion verification
  • MCP read-write architecture models and agent-maintained documentation
  • coding harness governance review gates and change management

Measured heat

now 0 pts/hpeak 0 pts/hcomments 0/hpeers p14momentum: steady2 platformsage 718h
points/hour across evidence · reading as of 2026-10-12 02:59:37.977291+11:00 · deterministic, not a model opinion

How the heat travelled

09-11 18:40 (minted)⭐ origin echo-reconstructedViaduct's product page offers a C4 architecture model readable and writable over MCP and describes change sets pinned to a fixed version, wi
Quiet Grid Labs on blog (echo) · attributed from hn.story.49662336 · published time unknown
—
09-11 17:43first on hacker news · published · lag ?Show HN: Viaduct – Architecture change sets for AI coding agents
igrlgkv
—
09-11 17:43amplified on hacker news 👑hn.story.49662336
igrlgkv
peak 3 · 0 comments · 43% of case engagement
09-15 04:11amplified on hacker newshn.story.49707601
ake2l
peak 1 · 0 comments · 15% of case engagement
09-24 05:24amplified on hacker newshn.story.49826584
ake2l
peak 3 · 0 comments · 43% of case engagement
09-11 18:21our radar first saw it · lag ?discovery anchor: hn.story.49662336—
pace: p42 vs 1032 stories at the 336h mark (now 718h old) — ahead of agent-memory-add-search-evaluation (1.2x), behind agenticos-self-hosted-governance (0.9x)

Evidence (4) — ⭐ canonical anchor

sourceobjectauthorscorecomments
🟧 hnShow HN: Viaduct – Architecture change sets for AI coding agents
Retrieved article excerpt

Open article · Retrieved 2026-09-11T18:22:56.091396+00:00

C4 modelling · MCP server · self-hosted A C4 model your team maintains and your agents can read Draw system context, containers, components and code. Attach the docs, sequence diagrams and API contracts to the elements they belong to. Serve the whole thing to Cursor or Claude Code over MCP. Quick start Log in free · no account · runs in this browser · export any time 01 Levels Four levels, one model Open a system to see its containers, a container to see its components. Every level is a view of the same model, so a rename or a new dependency shows up everywhere it belongs. 02 Notation The technology picks the colour Pick a technology and the element takes its colour and logo. Datastores are drawn as cylinders, people as actors — so a diagram is scannable before anyone reads a label. 03 What an element carries Docs, diagrams and schemas live on the element Documentation Sequence diagrams ER schema & API Every element carries its own Markdown pages — a design note on a service, a runbook on a worker, an ADR on a system. Write in the editor with a live preview, insert a stored sequence diagram or an API endpoint as a block, and the page keeps rendering the current version instead of a screenshot that rotted last quarter. 04 Testimonials What teams say once the model is theirs The tool earns its place when the diagram stops being a screenshot someone has to remember to update. I built Viaduct because I kept redrawing the same diagram to answer the same question. Now the model answers it, and my agent reads it too. IG Igor Golovko Founder, Quiet Grid Labs The C4 levels finally match how we actually talk about the system. Drilling from context into containers takes one click instead of three diagrams. MK Marta Kowalczyk Backend lead, fintech Our architecture diagram used to be a screenshot in Confluence, six months stale. Now it is the thing we open during incident review. DO Daniel Osei SRE, marketplace platform The MCP server is the part that sold the team. Cursor answers questions about our services without anyone pasting a diagram into the chat. PR Priya Raghunathan Staff engineer, logistics I built Viaduct because I kept redrawing the same diagram to answer the same question. Now the model answers it, and my agent reads it too. IG Igor Golovko Founder, Quiet Grid Labs The C4 levels finally match how we actually talk about the system. Drilling from context into containers takes one click instead of three diagrams. MK Marta Kowalczyk Backend lead, fintech Our architecture diagram used to be a screenshot in Confluence, six months stale. Now it is the thing we open during incident review. DO Daniel Osei SRE, marketplace platform The MCP server is the part that sold the team. Cursor answers questions about our services without anyone pasting a diagram into the chat. PR Priya Raghunathan Staff engineer, logistics Change sets let us review an architecture change the way we review code. The diff is the conversation. TL Tomas Lindqvist Head of platform, SaaS Sequence diagrams next to the model, not in a separate tool that nobody updates. That alone saved us the weekly sync. ED Elena Duarte Tech lead, healthtech We imported an OpenAPI spec and had the endpoints on the container in a minute. No one had to draw a box by hand. AB Ahmed Barakat Integration engineer, telecom New joiners read the model on day one and ask better questions on day two. HB Hannah Brecht Engineering manager, e-commerce Change sets let us review an architecture change the way we review code. The diff is the conversation. TL Tomas Lindqvist Head of platform, SaaS Sequence diagrams next to the model, not in a separate tool that nobody updates. That alone saved us the weekly sync. ED Elena Duarte Tech lead, healthtech We imported an OpenAPI spec and had the endpoints on the container in a minute. No one had to draw a box by hand. AB Ahmed Barakat Integration engineer, telecom New joiners read the model on day one and ask better questions on day two. HB Hannah Brecht Engineering manager, e-commerce 05 Magic flows A request, played back across the canvas A Magic flow is an ordered walk through the model: each step names a hop, the elements it goes from and to, and the connection it uses. Steps that happen at the same time share a stage, so a fan-out reads as a fan-out. — Build a flow from the elements you already modelled — no second diagram to maintain. — Play it on the canvas: everything dims except the current hop, and the camera follows the request. — Generate a sequence diagram from the flow, and attach the docs and diagrams that explain each step. Playing “Payment leaves the bank”, step by step. 06 Contracts Endpoints are stored as OpenAPI Describe an endpoint in the editor — typed parameters, request body, a card per status — and the container holds the OpenAPI document for the whole service. Import a spec to fill the canvas from it; export it back as a file. { "openapi" : "3.1.0" , "info" : { "title" : "Accounts API" , "version" : "1.4.0" } , "paths" : { "/api/accounts/{id}" : { "get" : { "operationId" : "getAccount" , "summary" : "Get account" , "parameters" : [ { "name" : "id" , "in" : "path" , "required" : true , "schema" : { "type" : "string" } , "example" : "acc_42" } ] , "responses" : { "200" : { "description" : "OK" , "content" : { "application/json" : { "example" : { "id" : "acc_42" , "currency" : "EUR" } } } } , "404" : { "description" : "No account with that id" } } } } } } accounts-api-openapi.json — exported from a container 07 MCP One server, one token, the whole architecture Cursor, Claude Code, Claude Desktop and any other MCP client can read the architecture straight from Viaduct: systems and services, API endpoints with their contracts, WebSocket channels, documentation and sequence diagrams. One hosted MCP server, one personal token, no repository to clone. — Your agent answers “what talks to what, and over which contract?” from the model instead of guessing from the codebase. — It can write back too — new services, endpoints, connections and docs land in the same model your team reads. — Hand it a change set: a piece of work pinned to a fixed version, with the constraints and acceptance criteria it has to prove. How the MCP server and change sets work Work an agent can finish, and you can verify A change set is scoped to one pinned version, so nothing shifts under the agent while it builds. It reports each finished piece with the criterion it proves and the commit that proves it — the status moves itself from draft to done, live on your screen. 08 Figma Your designers already use Figma. Now the model does. Connect a Figma account in settings and the frames your interface is drawn in become part of the architecture: components point at real nodes, and the values they are allowed to use are read from the file rather than retyped. Nothing is written back, and no picture is copied here. It tells you when the design moved Work is pinned to the file as it was. When somebody edits the frame afterwards, the change set says which components are now behind — instead of that turning up in review. One palette, not forty hexes The colours, spacing and type in the file come in by name. A component may use those, and a value that is not among them is flagged rather than quietly shipped. Your coding agent gets it too Everything here rides along in what an agent reads over MCP — the frame, its states, the palette. It builds the loading and error states because they were written down, not because someone remembered. 09 Assistant The boring half of documenting, done for you Not every team keeps an agent of their own, and the parts of a model that go undocumented are the dull ones: descriptions nobody wrote, flow steps still called "Step 4". Viaduct's built-in assistant writes those. It is in the editor, it needs no key and no setup, and it never changes your model on its own — every suggestion is shown, and applying it is a button you press. — A description for any element that lacks one — one at a time, or every blank in a system at once. — A review of the whole model: what is undescribed, what nothing talks to, what looks like the same job done twice. Every finding names a real element, and clicking it takes you there. — Names for the steps of a Magic flow that nobody got round to naming — written into the draft, so saving is accepting. Included with an account, with a monthly allowance per person. Guests draw; the assistant needs a sign-in. A review, as it comes back High Order Service No description — nothing says what it is responsible for, and three services depend on it. Medium Notification Worker Nothing connects to it and it connects to nothing; either the flow is missing or the service is. Low Legacy Billing Appears to do the same job as Billing Service — one of the two is probably the old one. 10 Your data Your architecture, on terms you can check An architecture model is a map of everything you run, which makes “where does it live?” the first fair question about a tool like this. Three answers, all of them verifiable. It stays where you put it Without an account the model never leaves your browser. With one it lives on servers in the Netherlands, inside the EU, and so do the backups. You can take it with you The whole model exports as JSON, and a container’s endpoints as an OpenAPI document — any time, without asking us. Nothing here is a one-way door. The agent gets in only if you let it MCP access needs a personal token you issue yourself, scoped to your projects. Revoke it and the access ends that moment. Read the privacy policy 11 Questions Answered before you have to ask What does it cost? Nothing. Viaduct is free — every level, the MCP server, collaboration and change sets included. There is no trial that ends and no feature held back behind a plan. If you want the work to continue you can donate, and that buys nothing but the work continuing. What is the AI assistant, and what does it cost me? It writes the parts of a model people leave blank: descriptions for elements that have none, names for unnamed flow steps, and a review of the whole model that points at real elements rather than giving generic advice. It runs on our provider, so there is no key to obtain and nothing to configure, and every account gets a monthly allowance of it. Nothing it suggests is applied on its own — you see the text and press a button, or you do not. Do I need an account? Not to draw. The editor opens with a sample model and works entirely in your browser, where the model stays on your machine and never reaches our servers. An account is for the things a browser cannot do alone: keeping a project across devices, sharing it with your team, and serving it over MCP. Where does my architecture live? On servers inside the EU, along with the backups. Without an account it lives in your browser and nowhere else. Either way you can export the whole model as JSON, and any container’s contracts as an OpenAPI document, whenever you like. What can an AI agent do with it? Read the model and write to it: Cursor, Claude Code, Claude Desktop or any other MCP client can ask what talks to what and over which contract, and can add services, endpoints, connections and documentation back. Access needs a personal token you issue yourself, scoped to your projects, and revoking it ends the access immediately. How is this different from a diagramming tool? A diagram is a picture; this is a model. An element exists once and appears on every level it belongs to, so a rename or a new dependency shows up everywhere at once. The endpoints on a container are a real OpenAPI document, not labels on a box — which is what makes the whole thing readable by something other than a person. Can I bring what I already have? Import an OpenAPI specification and the endpoints land on the container with their parameters, request bodies and responses. Models export and import as JSON, so moving a workspace — or keeping a copy outside this service — is a
igrlgkv30
🟧 echo.blog ⭐Viaduct's product page offers a C4 architecture model readable and writable over MCP and describes change sets pinned to a fixed version, wiQuiet Grid Labs——
🟧 hnArchkeel – coding agents declare architecture changes before they submitake2l10
🟧 hnArchKeel – results from 2 large agentic refactoring experimentsake2l30

Interpretation history

Decision trace