T

Stéphane Tailland 37770f39d4 feat(mcp): cap read volume per caller identity

A scope authorises a read and says nothing about volume. An agent reads a
thousand records in seconds where a person clicks ten, and until now nothing
noticed: the token is valid, the tool is permitted, the ACL passes. That gap
stopped being theoretical the day an autonomous run started reading the whole
catalog on a schedule.

identifier: credential keys the counter on the sub claim — the end user,
whoever is asking on their behalf — so an agent looping under one person's
identity spends that person's budget and cannot spend everyone's. Verified on
catalog: pm-user and ro-user hold separate counters.

Per-route namespace, so a busy domain cannot exhaust another's.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

2026-08-29 18:05:45 +02:00

dev

feat(mcp): cap read volume per caller identity

2026-08-29 18:04:14 +02:00

kong

feat(mcp): cap read volume per caller identity

2026-08-29 18:05:45 +02:00

.gitignore

First commit

2026-08-20 13:24:10 +02:00

openapi.yaml

Regenerate the spec against the gateway, not the backend

2026-08-27 06:42:02 +02:00

README.md

MCP server

2026-08-20 17:27:09 +02:00

README.md

catalog-api

Kong routing/policy layer for the catalog slice of classic-models-api (see /Users/stephanetailland/git/classic-models-api), split out per the DDD decomposition in ../kong-konnect-deployment-strategy.md. The backend itself stays a single monolith — only Kong does the domain split.

Two exposure surfaces of the same API, both proxying to the real classic-models-api backend:

REST (kong/rest/) — 14 routes covering products/productlines.
MCP (kong/mcp/) — the same catalog exposed as MCP tools (list-products, get-product, create-product so far; extending to the rest is mechanical), via ai-mcp-proxy.

Auth

openid-connect on both surfaces, with a distinct audience each (catalog-api for REST, mcp-catalog for MCP) and a real RFC 8693 token exchange chain: MCP → catalog-api → classic-models-api (REST's own exchange target), narrowing the token before it ever reaches the backend — the backend never sees a token minted for a different audience. consumer_claims/consumer_by map the JWT sub to a Kong Consumer for per-consumer rate limiting; consumer_groups_claim maps realm roles to Kong Consumer Groups for MCP's per-tool ACLs.

Other policies

Per-consumer + API-level rate limiting, response caching (GET/HEAD), audit log (http-log, endpoint still a placeholder — no real log collector in this lab).

Structure

dev/gitrepository.yaml + dev/flux-kustomization.yaml — this repo, self-referential; deploys ./kong into the catalog domain's dev environment (dependsOn the domain's Gateway infra in dataplane-catalog).
dev/externalsecret.yaml — git deploy key (ESO/Vault).
dev/token-exchange-broker-externalsecret.yaml — the shared client secret Kong uses to authenticate its own token-exchange calls to Keycloak.
kong/rest/, kong/mcp/ — see above.
openapi.yaml — extracted from the live backend, filtered to this domain's paths/schemas.