The Game Pensieve MCP
Click on "Install Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@The Game Pensieve MCPSearch my video games for RPGs released after 2015"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
The Game Pensieve — MCP Sidecar
A read-only MCP (Model Context Protocol) server for The Game Pensieve API. It is a sidecar proxy: a separate TypeScript process that exposes MCP tools over Streamable HTTP and fulfills them by calling the existing REST API over HTTP. AI hosts (Claude Desktop, Claude Code, claude.ai connectors) connect to it to answer natural-language questions about a game collection.
The API it proxies lives in a separate repo, the-game-pensieve-api; the references below to the
compose files (dockerCompose/compose.unsecured.yaml / dockerCompose/compose.secured.yaml for development,
dockerCompose/compose.production.yaml for production), Caddyfile, keycloak/, and the realm import all point
into that repo.
OAuth
When enforcing, the sidecar validates the incoming Authorization: Bearer JWT (signature via JWKS —
RS256 only, pinned explicitly — plus iss, aud, and the required scopes) with
jose, publishes Protected Resource Metadata
(RFC 9728) at /.well-known/oauth-protected-resource[/mcp], and challenges missing/invalid tokens
with 401 + WWW-Authenticate: Bearer resource_metadata="…". Valid tokens are forwarded to the API,
which independently validates them and scopes the request to the token owner (Keycloak sub → owner
→ RLS). Audience (aud) is validated on both sides — the sidecar and the API — to block confused-deputy
replay even though the API is on a private network.
Enforcement is gated by MCP_AUTH_MODE:
Mode | Behavior |
| enforce iff the backend heartbeat reports |
| always enforce (the recommended prod setting — no probe dependency) |
| never enforce (tokenless) |
In auto mode the startup heartbeat probe is retried (MCP_HEARTBEAT_RETRIES × MCP_HEARTBEAT_RETRY_DELAY_MS)
so a sidecar that boots before the backend is ready doesn't latch enforcement off from a single failed
probe. If the backend is still unreachable after the retries and OAuth is configured, the sidecar
fails closed (enforces) rather than serving /mcp tokenless.
iss is validated against the canonical, host-facing issuer, while keys are fetched from
MCP_OAUTH_JWKS_URI — the internal keycloak:8080 URL on the docker-compose network (in prod both are the
public https://… URLs). /healthz and the metadata endpoints stay public.
Scope enforcement
A token that verifies is not automatically authorized. After jwtVerify the sidecar requires every scope
in MCP_OAUTH_REQUIRED_SCOPES (defaulting to whatever MCP_OAUTH_SCOPES advertises) to be present in the
token's space-delimited scope claim. A verified but under-scoped token gets:
HTTP/1.1 403 Forbidden
WWW-Authenticate: Bearer error="insufficient_scope",
error_description="The access token is missing the required scope(s): pensieve:read",
scope="pensieve:read", resource_metadata="https://…/.well-known/oauth-protected-resource/mcp"403 rather than 401, per RFC 6750 §3.1: the credential is genuine, so re-presenting it would fail
identically, and a client that reads 401 as "start the OAuth flow" would loop. The scope parameter tells
the client what to ask for next; resource_metadata still lets it rediscover the authorization server.
Why this is not redundant with the audience check: Keycloak does not honor the RFC 8707 resource
parameter, so the /mcp audience is attached by a mapper on the pensieve:read client scope. In the
production realm that scope is deliberately not a realm default, so audience and scope now travel together
and are checked together. Set MCP_OAUTH_REQUIRED_SCOPES="" to disable the check — the escape hatch when a
connector cannot be granted the scope in Keycloak — which leaves audience as the only gate.
Related MCP server: GPT MCP Service
Tools
Tool | REST endpoint | Purpose |
|
| Valid fields + operators for an entity |
|
| Filtered search; the entity |
|
| Custom field definitions |
|
| Item counts per entity + total (computed server-side, no row transfer) |
|
| Public showcases (slug + name) |
All tools are read-only. Filters are { field, operator, operand } objects; call
get_available_filters first to learn the valid combinations. An empty/omitted filter list returns
everything.
Configuration
Env var | Default | Meaning |
|
| Backend base URL incl. |
|
| Port the |
|
| Per-request timeout to the backend |
|
|
|
| — | Expected token |
| — | JWKS URL for verification (compose: |
| — | Expected token |
|
| Scopes advertised in protected-resource metadata |
| (= | Scopes a verified token must carry to reach |
|
|
|
|
| Delay between heartbeat probe attempts |
Develop
Requires Node 20+. All configuration is via environment variables — see the table above and the
annotated .env.example.
npm install
npm run dev # tsx watch, reads .env-style vars from the environment
npm test # vitest (hermetic; no backend needed)
npm run typecheck
npm run build && npm startOn startup the server probes GET /v1/heartbeat and logs the backend's secureMode.
For architecture, request flow, testing patterns, and how to add a tool, see documentation/DevDocumentation.md.
Connect a host
Local (tokenless)
Bring up the API in permit-all mode (default local/docker profile) — no auth, the full read
surface is open:
API_BASE_URL=http://localhost:8080/v1 PORT=8090 npm start
npm run inspect # MCP Inspector — point it at http://localhost:8090/mcp
# or register with a host (no token needed):
claude mcp add --transport http pensieve http://localhost:8090/mcpMCP Inspector
MCP Inspector is the official debugging GUI for
MCP servers, installed here as a dev dependency (npm run inspect). It connects as a host would —
but with you driving instead of a model — which makes it the quickest way to see how the sidecar is
put together:
Tool discovery — the full tool list with names, descriptions, and Zod-derived input schemas, exactly as a host's model sees them.
Manual calls — invoke any tool from a form and inspect the result, including
isErrorresponses (e.g. the402/403capability responses surfaced from the backend).Raw wire traffic — every JSON-RPC request/response over the Streamable HTTP transport, useful when debugging the handshake or a host integration.
Select transport Streamable HTTP, URL http://localhost:8090/mcp. Against a secured backend it
prompts for the OAuth flow.
Secured (OAuth 2.1)
When the backend runs the secured profile the sidecar enforces OAuth. A host discovers the
authorization server from the sidecar's protected-resource metadata and runs the standard OAuth flow
(DCR + authorization-code + PKCE) against Keycloak; each user only sees their own collection.
Claude Code —
claude mcp add --transport http pensieve https://<MCP_DOMAIN>/mcp(locally,http://localhost:8090/mcpagainst the secured dev stack). On first use the CLI opens the browser to Keycloak to authorize; unauthenticated calls are challenged with401 + WWW-Authenticate.claude.ai / Claude Desktop connectors — add a Custom Connector with the remote MCP URL
https://<MCP_DOMAIN>/mcp. The client reads/.well-known/oauth-protected-resource, registers via DCR, and completes the PKCE login against Keycloak. (For remote hosts, pre-registering a client in Keycloak is more reliable than anonymous DCR — seekeycloak/README.mdin the API repo.)MCP Inspector — point it at the
/mcpURL; it will prompt for the OAuth flow.
The backend owner-scopes every tool call: a lapsed/guest caller gets the same 402/403
capability responses (surfaced as MCP isError results) it would from the REST API — MCP reads
inherit the exact same authorization as the web app (RLS + the capability matrix), never more.
Docker / Compose
The docker-compose stack lives in the API repo, which consumes this sidecar as a published image
(sethcondie/the-game-pensieve-mcp:latest) — the same way it consumes the front end. Build and push
from this repo:
docker build -t sethcondie/the-game-pensieve-mcp:latest .
# multiplatform build-and-push commands: documentation/DevDocumentation.md in the API repoThen, from the API repo:
docker compose -f dockerCompose/compose.unsecured.yaml up -d backend mcp
# MCP endpoint: http://localhost:8090/mcp
# Use -f dockerCompose/compose.secured.yaml instead to develop against a backend that requires tokens; the
# sidecar's auto mode reads the backend heartbeat and starts enforcing OAuth on its own.To iterate locally without a rebuild each time, run npm run dev on the host against the docker-compose
backend (API_BASE_URL=http://localhost:8080/v1 PORT=8090) instead of the mcp service.
In production (dockerCompose/compose.production.yaml in the API repo) the sidecar has no public ports — Caddy
fronts it at https://<MCP_DOMAIN>/mcp and terminates TLS. There it runs MCP_AUTH_MODE=required
with the public issuer/audience URLs. See that repo's Caddyfile + dockerCompose/.env.production.example.
Transport notes
Streamable HTTP, stateless: each POST /mcp spins up a fresh server and transport. GET/DELETE
on /mcp return 405 (no server-initiated streams or sessions). GET /healthz is a liveness probe.
This server cannot be installed
Maintenance
Resources
Unclaimed servers have limited discoverability.
Looking for Admin?
If you are the server author, to access and configure the admin panel.
Related MCP Servers
- AlicenseAqualityAmaintenanceA read-only Model Context Protocol server for Valve's public Steam Web API and storefront. Ask about your friends and games, playtime and achievements, plus account-independent things like sales, reviews, live player counts, Steam Deck compatibility, discovery, and recommendations — bring your own free Steam API key.372MIT
- FlicenseNot gradedqualityBmaintenancePrivate OAuth-backed MCP server for ChatGPT, supporting GPT Apps via MCP Streamable HTTP and GPT Actions via REST endpoints with OpenAPI 3.1.
- AlicenseNot gradedqualityDmaintenanceA feature-rich Model Context Protocol server that gives AI assistants full access to your Steam game library.144MIT
- AlicenseNot gradedqualityBmaintenanceAn MCP server that exposes your Steam, Epic Games Store, and IGDB game data as tools, enabling game library queries, install status checks, and metadata enrichment.1MIT
Related MCP Connectors
MCP server for Argo RPG Platform — connects AI assistants to campaign data via OAuth2
Official remote MCP server for Archivist AI TTRPG campaign memory: characters, sessions, and more.
Agent-native MCP server over the public saagarpatel.dev corpus. Read-only, stateless.
Latest Blog Posts
- Who's Calling? MCP Hosts Are an Identity Blind Spot (And the Spec Knows It)By Om-Shree-0709 on .mcpAgent IdentityOAuth 2.1
- Your AI Chatbot Just Exposed Your CEO's Salary to an InternBy Om-Shree-0709 on .Agent IdentityMCP SecurityOAuth Delegation
- Why MCP Servers Need Execution Sandboxing (And Why Your Current Stack Isn't Enough)By Om-Shree-0709 on .Agentic AiPrompt InjectionWebAssembly
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/sethhaskellcondie/the-game-pensieve-mcp'
If you have feedback or need assistance with the MCP directory API, please join our Discord server