Archstone — connect your business to every AI

A compiler for AI capabilities. You describe what your business can do, once, in business terms. Archstone compiles that into tools an AI agent can discover and call — MCP today, other protocols as they arrive. Nobody hand-writes integration code.

Open source, Apache-2.0.

See it work — 60 seconds, nothing to install

A capability compiled by Archstone is running live. Point Claude at it:

https://demo.archstone.dev/mcp

Open Claude (web, desktop or mobile) → Settings (or Customize) → Connectors → Add custom connector → paste the URL. Then ask about a trip — a destination, dates, a budget. Works on Free too (one custom connector is all this needs).

Prefer a terminal:

claude mcp add --transport http archstone-tourism https://demo.archstone.dev/mcp

The backend behind it is a plain HTTP service, and you can curl it directly — the same data the agent sees, deterministic so the shape is obvious:

curl -s -X POST https://demo.archstone.dev/v1/search \
  -H "content-type: application/json" -d '{"destination":"Rome"}'

The entire integration that made this callable by an agent is 12 lines of business YAML — no HTTP, no JSON Schema, no MCP SDK. Everything else was generated.

Start from an API you already have

Point archstone init at an OpenAPI document. It reads the spec, asks you the questions no document can answer, runs the real compiler over what it drafted, and writes nothing at all if that does not compile.

archstone init openapi.yaml --out manifest --company acme --domain catalog

The one answer it never guesses is effect — read, write or irreversible is the difference between looking up a price and charging a card, and no spec says which. Where a response could honestly be read two ways, it asks rather than picking. With --probe it will also make one read-only call to your real backend and record a genuine fixture, so archstone verify has something true to replay later.

The spec in that recording is examples/demo/stays-openapi.yaml, describing the demo backend in this repository — you can run it yourself.

Who is running it

ArtVinci — a custom-framing business — answers customer questions today through a capability compiled by Archstone. Real catalog, real prices computed live by their own backend. See the case study.

Why a compiler, and not just an MCP server

Writing your first MCP server is not the hard part — it is a few hundred lines, and you can do it in an afternoon.

The work is the fifth one. ChatGPT, Gemini and whatever ships next each want the same capability shaped slightly differently, every one of them is a separate integration project, and your API keeps changing underneath all of them at once.

Archstone is not an MCP server. It is a compiler that, in its first release, generates one.

One capability definition (CDL) lowers to a target-agnostic IR; emitters consume the IR. MCP today; REST · GraphQL · SDK tomorrow. Change the protocol and you regenerate — you do not rewrite. Change the backend and the CDL and the generated tool do not move at all.

That is what zero manual integration means: not that the first server is easy, but that the maintenance disappears instead of multiplying.

How it works

capabilities.yaml   →   *.capability.yaml   →   bindings/*.binding.yaml
(what the company     (each capability:        (how one capability maps
 offers — the index)   business shape only)     to a real HTTP endpoint)

        └──────── archstone apply ────────┘        └── archstone serve ──┘
             parse → validate → compile → IR          emit MCP tools → agent

You describe capabilities in CDL (Capability Definition Language) — business only, no integration code. Archstone compiles that to a target-agnostic IR, and an emitter turns the IR into tools an AI agent can call. Swap the backend; the CDL and the generated tool do not change.

Capability outputs reference named resources (*.resource.yaml); the compiler resolves them into a typed, described outputSchema, and a binding's response: mapping enforces that shape at every call — a required field missing from the provider's response fails closed (a structured error, never a silent raw pass-through). archstone verify replays a recorded fixture against the live backend on demand and reports a 🟢/🟡/🔴 health status per binding, so contract drift shows up before an agent hits it — naming the fields the provider gained, lost or retyped, not merely reporting that something moved. Bindings whose capability effect is not read are skipped by default (replaying a fixture is a real invocation), and re-included only with --sandbox, an assertion that the backend is a sandbox tenant.

A field your manifest does not name never reaches a model. That is deliberate: your provider's payload very likely carries wholesale rates, commissions or internal ids beside the fields you publish, and a backend deploy adding one must not be a decision about what an assistant can say. Declaring a new field is a separate, deliberate act — archstone adopt offers each one, asks you to describe it, writes it into your resource and binding, and recompiles before keeping anything. With stdin closed it refuses and writes nothing: it needs a person, which is the point rather than a limitation.

Quick start

From source (this repository):

pnpm install

# Scaffold a manifest from an API you already have (opt-in, read-only, no LLM)
pnpm exec tsx packages/cli/src/index.ts init path/to/openapi.yaml --out my-manifest --domain catalog

# Compile a manifest: validate + lower to IR
pnpm apply examples/manifests/booking

# Build a portable IR artifact (for embedding in your own app)
pnpm build examples/manifests/tourism

# Serve it to an AI agent as MCP tools over stdio
pnpm serve examples/manifests/tourism

# Serve it as MCP over HTTP (e.g. for Claude API mcp_servers)
pnpm serve --http examples/manifests/tourism --token my-bearer-token

# Replay a binding's golden fixture against the live backend; detect drift
pnpm verify examples/manifests/tourism

# Replay including write/irreversible capabilities (only for sandbox backends)
pnpm verify examples/manifests/tourism --sandbox

# Get structured JSON output for integration with CI pipelines and dashboards
pnpm verify examples/manifests/tourism --json

From npm (standalone CLI):

# Install globally or use npx
npm install -g @archstone/cli
# or
npx @archstone/cli apply <manifest-dir>

# Then run the same commands:
archstone init path/to/openapi.yaml --out my-manifest --domain catalog
archstone apply examples/manifests/booking
archstone build examples/manifests/tourism
archstone serve examples/manifests/tourism
archstone serve --http examples/manifests/tourism --token my-bearer-token
archstone verify examples/manifests/tourism
archstone verify examples/manifests/tourism --json

# Check a manifest against the pre-production checklist — offline, no backend contacted
archstone doctor examples/manifests/tourism

# Declare a field the backend started returning (asks before writing; needs a person)
archstone adopt examples/manifests/tourism

Where your CDL lives

Your business's CDL manifest (capabilities.yaml, *.capability.yaml, *.resource.yaml, and bindings/*.binding.yaml) is authored and version-controlled in your own application repository — never inside this Archstone repository or any other Archstone-owned repository.

@archstone/cli is a stateless compiler. It runs locally on your machine or in your own CI pipeline with zero checkout of any Archstone repository required — public or private. Install @archstone/cli from npm; point it at your manifest directory; it compiles to IR and reports the result. That's the entire integration: no cross-repo credentials, no monorepo dependency, no fetch-at-runtime.

Distinguishing "From source" above: the instructions above for exploring Archstone's source code are for contributors building Archstone itself. The real integration path for your business is to install @archstone/cli from npm into your own repository and wire archstone apply/archstone build/archstone serve/archstone verify into your own build system. See the onboarding guide for the full walkthrough.

New here? Start with the onboarding guide — one path for providers (expose your business to agents) and one for contributors (build Archstone).

Embedding Archstone

Rather than running archstone as a separate CLI or MCP server, you can embed the compiled IR directly in your own agent loop. After building a portable IR with archstone build, consumers can use the @archstone/agent SDK (RFC-0008):

import { fromIR, tools, execute } from "@archstone/agent";
const archstone = fromIR(compiledIR);

// Get typed tool definitions in your preferred format
const myTools = archstone.tools("anthropic"); // or "openai" / "gemini" / "json-schema"

// Invoke capabilities directly — no MCP server process needed
// Accepts both raw dotted id and sanitized tool name (as returned by tools())
const result = await archstone.execute("tourism.search", { location: "Paris" });
// or: await archstone.execute("tourism_search", { location: "Paris" });

For those who want HTTP-based MCP (e.g., to expose an embedded instance via Claude API's mcp_servers), the /mcp subpath provides a mountable Streamable-HTTP handler:

Read the full README on GitHub →