Skip to main content
Glama

a2a2p — Agent-to-Agent-to-Physical

An open protocol for turning intent into physical things.

An AI agent can reason about a bracket. It cannot make one exist. a2a2p is the missing layer: describe what needs to be true about the world, and get back a manufacturable specification, an engineering review grounded in real material properties and real geometry, and a price.

The specification is public domain (CC0) — implement it freely, including in products competing with this one. a2a2p.com is a reference implementation, not the protocol.

Abundance arrives as the distance between knowing what is needed and the thing existing collapses toward zero. Today that distance is mostly friction. This is an attempt to remove it.


Try it in 30 seconds

No API key. No signup. No auth.

claude mcp add --transport http a2a2p https://a2a2p.com/mcp

Or any MCP client:

{ "mcpServers": { "a2a2p": { "type": "http", "url": "https://a2a2p.com/mcp" } } }

Client only speaks stdio? Bridge to it:

npx a2a2p-bridge                      # or: docker run -i --rm ghcr.io/deadlypillow/a2a2p

The bridge is dependency-free and takes A2A2P_ENDPOINT, so it works against any implementation of the spec, not just this one. The hosted endpoint is dual-era: finalized MCP 2026-07-28 clients use stateless server/discover and per-request metadata, while existing 2025-03-26 clients keep the legacy initialize path.

Or plain HTTP:

curl -X POST https://a2a2p.com/api/review \
  -H "Content-Type: application/json" \
  -d '{"intent":{"purpose":"Outdoor bracket holding a 10kg sensor in direct sun","environment":["outdoor","uv_exposure"],"quantity":20},
       "specification":{"process":"fdm_3d_print","material":"pla","tolerance_class":"precision"}}'

You get back, instantly: PLA degrades in UV — specify a UV-stable grade or switch material. As-printed accuracy is ±0.2–0.5 mm; "precision" needs post-machining or a relaxed class. Plus a quote-readiness score and what to add to raise it. Nothing is stored.


Related MCP server: Fabric Marketplace

Capabilities

Tool

What it does

Latency

review_specification

Engineering review with auditable beam deflection/stress for explicit load cases, a compact next patch, visible completion path, optional CAD accelerators, and a direct next_call MCP/REST handoff — never invented defaults. Returns readiness on two axes: resolution_readiness (what you owe) and quote_readiness (what a supplier needs).

instant, stateless

get_supply_options

Deterministic price and lead-time estimates today; pricing_kind distinguishes model estimates, provider estimates, and binding provider quotes. Attach an STL (measured fully) or a STEP file (envelope, context-resolved units, machining complexity, and the CAD system's declared solid volume where present). Every geometry figure names its provenance.

instant

request_physical_solution

Full decision-support resolution: recommended path, buy-vs-make, and delivery risks. Returns immediately with the review embedded.

sub-second

revise_requirement

Apply a sparse clarification patch to a prior request without losing its decision trail. Creates an immutable linked revision with a deterministic before/after readiness comparison; supports retry idempotency.

sub-second

check_request_status

Retrieve the resolution report and any persisted revision-readiness feedback.

instant

Also served: /spec · /conformance · /policy · /api/providers · /evidence/providers · /llms.txt · /schema

/api/providers publishes a confirmed status, the raw most recent probe, and a measured 24-hour reliability window (samples, availability, probe latency, last failure) per spec §5.7. A transient probe failure is retried once and reported as degraded until a second consecutive failure confirms it — one 15-second timeout should never be able to publish down for a provider that is answering everyone else. Availability is a2a2p's own reachability measurement, not a provider service-level commitment and not fulfilment capability.


The idea worth stealing

A physical requirement has two layers, and conflating them is the main reason manufacturing communication fails:

  • Intentwhat must be true about the world. "Holds 15 kg, outdoors, ten years."

  • Specificationwhat object satisfies that. "6061-T6, 4 mm, anodized."

Most systems demand the second and discard the first. Then, downstream, someone proposes a substitution and nobody knows what the part was for. a2a2p accepts intent alone as a complete request — deriving the specification is the protocol's central act of value — and carries the intent forward permanently.

Two rules in the spec are normative, and they are worth adopting even if you use none of this code:

  1. An estimate MUST NOT be presented as a quote (§5.2.1). Every price is labeled model_estimate, provider_estimate, or provider_quote; only the last is a supplier commitment. An agent that cannot tell these apart will eventually commit real money to a fiction.

  2. Anything that fabricates MUST screen requests and publish its policy (§5.5). Silent refusal is non-conformant — an agent that cannot learn why it was refused cannot correct course.

  3. Don't blame the requester for what you promised to derive (§5.1.1). If you derive material, process, dimensions and tolerance from intent, listing them as the requester's missing fields tells an agent to go and find the information you just said you would produce. Report readiness per stage: have you told me enough? and is this ready for a supplier? are different questions, and one blended score answers neither.


Conformance

L1 SpeaksL2 PricesL3 ResolvesL4 Fabricates. Self-declared; there is deliberately no certifying authority.

Verify any implementation yourself — the suite ships with the spec:

node tests/conformance.mjs https://a2a2p.com     # a2a2p semantics
node tests/mcp-2026-conformance.mjs https://a2a2p.com # MCP 2026 binding

The MCP probe checks stateless discovery, standardized HTTP headers, cache hints, structured tool results, error semantics, and the retained legacy initialize path. Its calls are marked probes and do not create durable requests.

The semantic suite reports its current assertion count at runtime, with each assertion citing the spec section it enforces. It deliberately cannot prove fabrication (that would mean placing real orders), so it verifies the L4 prerequisites and records the fulfilment claim as unproven — otherwise any implementation could publish a policy, make nothing, and pass as L4.

This implementation declares L3 and publishes exactly where it falls short at /conformance — including that L4 is unmet, that STEP solid volume is the CAD system's declared property rather than a B-rep evaluation, and that deterministic screening cannot read intent. An implementation that will not hold itself to its own spec has written a brochure.


Honest status

Built deliberately ahead of demand. As of August 2026:

  • ✅ Assessment, bounded beam physics, pricing, geometry measurement, screening, and resolution reports are live and free.

  • ✅ The first authenticated provider-generated estimate is preserved as public evidence: a screened PLA sample returned $1.40 from Slant 3D. It was not a quote, order, payment, fabrication request, shipment, or delivery.

  • ⚠️ No fulfillment path is active. The implementation is L3: it does not route supplier RFQs, auctions, payment, orders, or fabrication. Any L4 activation must meet the evidence gate in docs/FULFILLMENT_ACTIVATION.md.

  • 📐 STL meshes are measured fully (volume, area, complexity) and carry volume_source: measured_mesh. STEP files yield envelope, feature-based machining complexity, and units resolved from the shape's own representation context — reading an inch file as millimetres is a 16,387× error on volume, so it is resolved, never sniffed. Where the exporter wrote a geometric validation property, the CAD system's own solid volume is read, unit-converted, checked against the measured bounding box, and labelled volume_source: declared_validation_property — a2a2p did not compute it. Assembly component properties are never summed into a part volume, and a2a2p does not evaluate trimmed NURBS.

If you are a manufacturer with a programmatic quoting API, or building one: hello@a2a2p.com. That is the bottleneck for everyone in this space, not just for us.


Repository

SPEC.md                    the specification (CC0) — the actual product
src/worker.js              reference implementation (Cloudflare Worker)
src/schema.js              canonical requirement schema
src/review.js              deterministic engineering review engine
src/mechanics.js           bounded beam deflection/stress sanity checks
src/geometry.js            STL mesh measurement (volume, area, complexity)
src/step.js                STEP entity-graph parsing (envelope, unit resolution, features, CAD validation properties)
src/supply.js              provider adapters + parametric estimator
src/screening.js           intake screening (the governance rail)
src/file-inspection.js     declared-vs-actual CAD format and bounded metadata checks
src/telemetry.js           demand instrumentation
bridge/index.js            stdio <-> Streamable HTTP bridge (dependency-free)
Dockerfile                 runs the bridge; for clients and registries that run servers
tests/test-worker.mjs      regression suite
tests/calibration.mjs      pricing model vs real published prices
tests/conformance.mjs      spec conformance suite — runs against ANY implementation
docs/                      research, listings kit, activation gate
PROTOCOL_LOG.md            build log — every session, decision, and failure

Run the tests: node tests/test-worker.mjs · Deploy: npx wrangler deploy

Contributing

Implementations of the spec, conformance reports, corrections, and competing implementations are all welcome — that is what a standard is for. Issues and PRs here, or hello@a2a2p.com.

Licensing, in two parts (see LICENSE):

  • Specification (SPEC.md) — CC0 1.0, public domain. Implement it freely, including in competing products, without permission or attribution.

  • Implementation (everything else) — MIT.


a2a2p — when agents need things to exist.

F
license - not found
-
quality - not tested
A
maintenance

Maintenance

Maintainers
Response time
Release cycle
Releases (12mo)
Commit activity

Related MCP Servers

  • F
    license
    -
    quality
    D
    maintenance
    An agent-native marketplace API where any agent can publish allocatable resources, search for what they need, negotiate structured offers, and exchange contact details after mutual acceptance. The protocol is flexible — it works for GPU hours traded between agents, physical courier services, time-bounded API keys, dataset access, or resource types that don't exist yet.
    Last updated
    1
  • A
    license
    -
    quality
    B
    maintenance
    AAAA-Nexus MCP provides 135 tools for autonomous AI agent infrastructure: formally bounded trust scoring, RatchetGate session security, on-chain A2A escrow, EU AI Act / NIST compliance certification, DeFi parameter proofs, VRF gaming with cryptographic proof, and verifiable randomness — all backed by the atomadic.tech API. Free tier available, no signup required.
    Last updated
    MIT

View all related MCP servers

Related MCP Connectors

View all MCP Connectors

Latest Blog Posts

MCP directory API

We provide all the information about MCP servers via our MCP API.

curl -X GET 'https://glama.ai/api/mcp/v1/servers/Deadlypillow/a2a2p'

If you have feedback or need assistance with the MCP directory API, please join our Discord server