Open source · .NET 11 · Angular 22 · Aspire
Tendero is an ecommerce platform for people and AI agents — a real shop with a real checkout, open to agents through WebMCP, MCP and UCP, and built in the open with an AI coding agent, one measured step at a time.
MIT licensed. Clone it, run one command, and it works — no sign-ups, no API keys, no cost.
Run it locally
Postgres, Elasticsearch, the API, the outbox worker and both Angular apps come up together under Aspire, with its dashboard in front. The AI stack — embeddings, vectors, a reranker — is self-hosted and starts on demand.
# .NET 11 preview SDK, Node 24, Docker
git clone https://github.com/juanluelguerre/tendero.git
cd tendero
(cd frontend && npm ci)
dotnet run --project src/AppHostWhy it exists
Tendero has to be two things at once — a credible commerce platform and an agent-native one — because an agent checkout over a pretend cart demonstrates nothing. Around that core it serves four more purposes.
01
Catalogue with variants and localized attributes, price lists and promotions with combination rules, real-time stock across warehouses, cart, checkout, taxes, shipping, payments, orders with a state machine, returns, accounts and guests, audit.
02
The same shop, open to AI agents through three surfaces with three trust models — and the aim of being the first open .NET reference implementation of the Universal Commerce Protocol.
03
Hybrid search with regression tests, a self-hosted AI stack that costs nothing to run, and a rule held by tests: every AI feature has a non-AI path, and the degradation is proved.
04
How to direct an agent without the project degenerating: a written constitution, architecture rules the agent cannot argue with, and gates that measure instead of trusting a confident answer.
05
The companion code to a blog series. Every article links a tag you can clone to get exactly the tree it describes — and a book on building Tendero is being written from the same notebook.
06
MIT licensed, with a hard rule for everything it links: permissive licences only, model weights included, each one checked at the exact version added.
Agent-native
They are not alternatives: each answers a different question. And an agent is a principal of its own, never a customer — which is why its actions are always shown in clay.
WebMCP
An agent in the shopper's own browser uses the shop's tools — search, add to cart — and inherits the shopper's session. No second principal, no token. Feature-detected: without it, the shop is unchanged.
Shipped
MCP server
Eight tools for any MCP client, over the same queries the shop itself runs. Anonymous by an explicit decision, and read-only by construction.
Shipped
UCP + AP2
A profile at /.well-known/ucp derived from the code, the catalogue over MCP and REST, and
validated against the protocol's own schemas. Checkout with signed AP2 mandates comes next.
Catalogue shipped · checkout in phase 11
Architecture
From outside, one system that people and agents use. Inside, ten containers under .NET Aspire. The drawing is held to the architecture model by a check in CI, so it cannot show a box the code does not run.
Shopper
Browses, fills a basket, pays
Agent in the tab
WebMCP · the shopper's session
MCP client
Reads the catalogue · anonymous
UCP agent
Own principal, signed mandate · phase 11
Shopkeeper
Reviews, prices, stock, orders
Tendero
One system: a storefront and a backoffice, an API with three agent surfaces, workers, a database and a search index — and an AI layer that starts on demand.
Tendero · .NET 11 and Angular 22 under Aspire
Storefront
Angular 22 · the shop
WebMCP tools
Backoffice
Angular 22 · review, orders, stock
API
.NET 11 · Carter slices over one command and query dispatcher
/api/mcp/.well-known/ucp
CatalogPricingInventoryOrderingAccounts
Workers
Drain the outbox: projections and the order saga
Postgres
Source of truth, outbox, audit
Elasticsearch
Lexical search · the permanent fallback
AI layer · on demand
Qdrant
Product vectors
Embeddings
bge-m3
Reranker
GPU only
Observability
Aspire dashboard always · the Grafana stack behind a switch
Identity provider
OIDC · a development issuer, or Keycloak
Payment provider
Behind a port · a fake adapter today
Drawn from the C4 model in the repository (docs/c4/), whose containers a check
holds to the AppHost's resources — and a check holds this drawing to the model.
What's different
Each one says where it stands today. Nothing here is claimed before it runs.
UCP capabilities, an MCP server and a browser agent surface, on the same application layer the shop uses.
Structured knowledge per product — specs, evidence, comparisons — each claim carrying its source and confidence. The AI proposes, the shopkeeper approves.
Identify legitimate agents — signatures, AP2-style mandates — and make anti-fraud rules agent-aware.
Natural language over the shop's own data and traces, proposing approvable commands — never prose, never a write on its own.
Constrained answers such as "a kit for under 100 €", showing their arithmetic.
A relevance gate on every pull request, one trace from the shopper's click to Elasticsearch, and next: cost per conversation and a trace for every decision.
How it is built
Architectural completeness is total; data volume is laboratory scale. And every rule below is held by a test, not by a comment.
A model never writes to the shop. It writes a proposal with its source and confidence; approval is a domain operation with an audit row.
Lexical search is the permanent fallback. When the vector layer is down, a search degrades byte for byte — and a test proves it.
Catalogue, Pricing, Inventory, Ordering and Accounts cross only with values and events. An order snapshots what it bought.
Catalogue sources, payments, identity, search, telemetry: a new connector is a class and a registration, and it passes the same suite as the others.
An order's ten transitions are declared in one table beside the aggregate, and nothing changes a status any other way.
The OpenAPI document is generated from the running API, held to it by a test, and generates the frontends' types.
Measured, not promised
An agent always sounds certain, so Tendero measures. These come from the repository's own gates.
Relevance is scored on golden sets of 52 queries per language across eight query types, and CI fails a pull request that regresses it. Mutation testing checks that the tests would notice.
Roadmap
Phases 0 to 9 are in. The code reaches the public repository with the article that explains it, so it is never ahead of its explanation.
Read how it is built, one article at a time, and clone the exact code each one describes.
The music is Tendero's own, composed for the film: warm while the shop has only people in it, gaining an electronic pulse when the agents walk in, and closing on its three-note identity. Hear the identity on its own · Download the film