the spec layer for agent-built software

Specs that can be graded.

Write the specification once. Point any agent at it to build. Then measure what got built — capability by capability, with evidence, pinned to a commit.

ProveSpec turns a specification into a checkable artifact: a versioned tree of capabilities that any implementation can be scored against. Ports, rewrites, agent output, competing builds — same measurement, called parity.

NANOCHAT · ENGINE LAYER · TARGET: .NET PORT 75% · 38 of 51 13 open implemented partial missing measured 2026-08-01 · evidence: source · commit-pinned
FIG. 1 — parity as a measured spanscope: AI (engine)
the exhibit

A real port, measured — not a mockup

Karpathy's nanochat (Python) is the reference. A .NET port chases it. Two agents extracted both specs from source; every grade carries a file reference. The report is what the porting agent works from — and it moves as work lands.

Engine scope · 51 capabilities

In-app Parquet ingestion — out-of-app JSONL conversion only
Distributed multi-GPU training — single-GPU only
Mixed precision (bf16/FP8) — fp32 only, ~2× KV memory
Optimized attention kernels — custom CUDA, no FA3
Tokenizer training — extra specials diverge vocab ids

Since this measurement, the porting agent closed several of these. That is the point: the number is a function of a commit.

The other direction

Parity is a directional measure, so the same tree answers the reverse question. The port has capabilities the reference never had — a desktop observatory, an MCP server, a model registry — which appear as target-only, not as defects.

And across layers

Grade a UI against its own engine: of 16 engine capabilities, 11 are reachable from the interface — 69%. Tokenizer training, the RL loop and the inference benchmark exist but no user can run them.

how it works

Five tools. Your agent does the rest.

STEP 01

Pull the spec

Your agent fetches a versioned capability tree — or checks a draft against the catalog to see what already exists.

STEP 02

Build against it

Any platform, any agent, any model. The spec is the contract; the environment is just a run parameter.

STEP 03

Grade what got built

Per capability: has it, partial, missing — with file-level evidence. Ungraded counts as missing, so the score can't be talked up. The gap list is the work queue.

# one line — no account, no install, no database
claude mcp add provespec -- npx -y provespec-mcp

# then, from any agent session:
spec_get(slug: "todomvc", scope: "AI")
  → the engine capability tree, ready to build against

spec_grade(slug: "todomvc", product: "my-build", grades: [...])
  → 75% · missing 6 · partial 7 · beyond-spec 4 · the gap list
the catalog

53 specifications · 13 categories · growing daily

Reference specs for things people actually build, plus extracted specs of real projects. Free to read, free to build against.

apps 11 services 7 engines 6 cli 4 formats 4 games 4 goods 3 hardware 3 libraries 3 protocols 3 layers 2 ports 2 generators 1

2048 · Auth Service · Calculator · Chat App · Coffee Grinder · Conway's Game of Life · CSV (RFC 4180) · Date & Time Library · Diff Engine · Excalidraw · Expense Tracker · Feature Flag Service · File Upload Service · HTTP Caching · HTTP Client CLI · Hugo · Job Queue · JSON Parser · JSON Tool CLI · JSON Web Token · Kanban Board · KV Store · Laptop · Markdown Editor · Markdown Parser · Minesweeper · nanochat (Python) · NanoChat.NET · NanoChat.NET — engine layer · NanoChat.NET — UI + MCP surface · OAuth 2.1 Authorization Code + PKCE · Obsidian (desktop) · Office Chair · PocketBase · Pomodoro Timer · ripgrep · RSS Reader · Running Shoe · Semantic Versioning 2.0.0 · Shell (TS workbench) · Smartphone · Snake · Solid-State Drive · SQLite · Template Engine · Todo CLI · TodoMVC · URL Shortener · Validation Library · Weather Dashboard · Webhook Relay · WebSocket · YAML 1.2

who it's for

Four questions, one measurement

porting or rewriting

"Is the rewrite actually done?"

Track a migration against its reference per commit. The gap list is the work queue; the trend is the status report. A CI check keeps it from regressing.

building with agents

"The agent says it's done — is it?"

Start from a spec instead of a prompt, then score what came back. Compare two builds objectively. Publish the badge.

evaluating agents & models

"Which run was actually better?"

One spec, many environments. Parity % as the score, the gap report as the diagnosis, evidence requirements to keep it honest.

maintaining a project

"What should someone build next?"

Publish your project's capability surface. Contributors see exactly which leaf is missing — and an engine with no interface is an open invitation.

pricing

Free where it grows. Paid where it works.

Public

Free

forever

  • Read every public spec
  • Submit assessments over MCP
  • Report links and README badges
  • Novelty check on a draft spec

Team

Paid

per team, monthly

  • Private catalogs and specs
  • CI parity gate on pull requests
  • History and trend dashboards
  • Higher submission limits

Platform

Talk to us

for agent and model builders

  • Benchmark sets and leaderboards
  • Verified audits by our own agent
  • Volume API access
  • Catalog licensing