About Proa documentation
Proa documentation is a read-only service exposing public Proa documents to agents and other clients. It provides search and paginated document retrieval through HTTP APIs and an MCP endpoint, without credentials or write access.
Key features
- 01Search public documentation
- 02Retrieve paginated documents
- 03Read-only access without credentials
- 04Markdown resources and tools
Use cases
- 01Searching Proa documentation
- 02Retrieving a specific documentation page
- 03Reading public docs through an MCP client
MCP support
The `/mcp` endpoint uses stateless Streamable HTTP and exposes two read-only tools plus paginated Markdown resources. It provides no write operations or credential access.
Configuration
Remote servers need no local install. Pick your client, then copy the config or install in one click.
https://proa.so/mcpclaude mcp add --transport http proa-web-docs https://proa.so/mcp{"mcpServers":{"proa-web-docs":{"type":"http","url":"https://proa.so/mcp"}}}Open in Cursor{"servers":{"proa-web-docs":{"type":"http","url":"https://proa.so/mcp"}}}Open in VS Code{"mcpServers":{"proa-web-docs":{"type":"http","url":"https://proa.so/mcp"}}}{"mcpServers":{"proa-web-docs":{"serverUrl":"https://proa.so/mcp"}}}Run commands in a terminal; paste JSON into the client’s MCP config and restart. Servers that need sign-in will prompt on first connect.
Tools
Details
From the project READMEProa Marketing
Rust SSR marketing site for Proa.
Agent interfaces
Authored document output uses typed Proa components and format-specific macros.
Each page owns its HTML and Markdown implementations and shares content fields
between them. The homepage composes explicit Markdown sections for the hero,
render examples, showcase comparison, benchmark charts, UI catalog, capabilities,
and quickstart. Markdown exposes all benchmark cases together and omits visual
demo fixtures and controls. Only the existing HTML-only highlighted code
fragments use the native converter. About/Contact pages also share their data;
authentication and the documentation index live in the docs module.
OpenAPI and MCP input schemas
use const_json! literals, while runtime discovery URLs use typed serializers.
Sitemaps use xml_sync!, and 404 pages use html_sync! / md_sync!.
The homepage supports Accept: text/markdown, /index.md, and ?mode=agent.
/llms.txt and /.well-known/llms.txt export that same page renderer; the
focused documentation index is /docs/llms.txt. Agent catalogs live at /.well-known/ard.json and
/.well-known/ai-catalog.json; the RFC 9727 API catalog is at
/.well-known/api-catalog. These documents use the resolved public site URL,
including Railway preview origins.
Public discovery files allow cross-origin browser reads and conditional-request
preflights. The /mcp transport retains separate Host and Origin checks. Its
experimental server card uses the same identity as the runtime handshake.
The skills index at /.well-known/agent-skills/index.json uses the draft
Agent Skills Discovery 0.2 format with SHA-256 digests for the served
SKILL.md files. These instructions describe workflows in a Proa source
checkout and reference its documentation and scripts.
/api/v1/search, /api/v1/docs, and /api/v1/docs/page provide typed,
paginated access to public documents. /mcp exposes the same source through
the official Rust MCP SDK's stateless Streamable HTTP transport, with two
read-only tools and paginated Markdown resources. /mcp/server-card describes
the connection. No credentials or writes are exposed. API and MCP responses
advertise an enforced shared per-instance budget of 600 requests per minute.
The optional /static/agent-tools.mjs registers the search and retrieval
operations with WebMCP. Its tests run with
node --test proa_marketing/tests-js/agent-tools.test.mjs. HTTP and MCP tests
run through the linked pipeline:
cargo proa test --root . --locked -p proa_marketing --bin proa_marketing
See the public API guide for the external contract. The deployment smoke check also verifies the discovery index, Markdown homepage, catalog, search response, and MCP server card. An Ora score only updates after deployment and a new external scan; search indexing and directory listings are outside the application's control.
Proa UI homepage rail
The homepage includes the complete home-rail-section from Proa UI's
marketing/src/pages/home.rs: ten inert component previews and the catalog
tile. Its rendered HTML and scoped CSS are checked in, so normal builds do not
need the downstream component library or another checkout. Caption links open
the live catalog at ui.proa.so.
To refresh it, run the Proa UI marketing application locally, then run this from the Proa workspace root with that application's URL:
node scripts/import-proa-ui-rail.mjs http://127.0.0.1:3333
The importer preserves the source rail's order, responsive styling, themes, and keyboard navigation. It removes hydration metadata from the inert previews and scopes their styles to this section. Review both themes and the mobile layout after refreshing the snapshots.
Docs search
The Cmd+K search runs entirely in the browser. The docs build emits
/docs/_search/documents.json (title, description, section, headings,
frontmatter keywords, and body text per page); public/search.mjs fetches it
on the first open and indexes it with a vendored Orama
bundle. Ranking lives in public/search-index.mjs, which the Node battery
imports too, so the score you measure is the score that ships:
pnpm run search:battery # local server on :5555
node scripts/search_battery.mjs --source prod
node scripts/search_battery.mjs --source path/to/documents.json --verbose
node scripts/search_battery.mjs --config tweak.json # override DEFAULT_CONFIG keys
The fixture is scripts/search-battery.json (one obviously-right page per
query, or a short list when two pages tie); its thresholds fail the run when
top-1 or top-3 drops. Give a page identifiers it canonically documents with a
keywords: [html_sync, html!] frontmatter list; that field is boosted above
body text. Regenerate the Orama bundle after bumping the @orama/*
devDependencies with pnpm run search:vendor.
The rsjs SearchModal island only owns the dialog (open, close, focus); see
the ownership note at the top of public/search.mjs.
Content
Marketing docs and blog content are compiled by proa_docs_build at build time.
Keep route-owned MDX in these crate-local directories:
src/docs->/docssrc/blog->/blog
Each collection follows the fumadocs-style meta.json convention for sidebar
ordering and nested groups.
Local Docker Build
Build from the repository root:
docker build \
-f proa_marketing/Dockerfile \
-t proa-marketing .
docker run --rm -p 5555:5555 \
-e HOST=0.0.0.0 \
-e PORT=5555 \
-e SITE_URL=http://localhost:5555 \
-e PROA_BUILD_REVISION=0123456789abcdef0123456789abcdef01234567 \
proa-marketing
Health check:
curl -fsS http://127.0.0.1:5555/healthz
Full public-contract smoke check:
python3 deploy/marketing_smoke.py \
--url http://127.0.0.1:5555 \
--expected-revision 0123456789abcdef0123456789abcdef01234567 \
--attempts 1
Docker reads .dockerignore from the build context root. This repository includes
a root .dockerignore so local target/, .tools/, and node_modules/
directories are not sent to the Docker daemon.
The Dockerfile builds entirely from this repository. No sibling checkout, additional BuildKit context, or registry credential is required.
The Docker builder regenerates proa_marketing/public/styles.css from
proa_marketing/styles/input.css using Tailwind's standalone CLI. No Node or
TypeScript workspace install is required. The build stage downloads the pinned
Tailwind 3.4.18 Linux executable for the target architecture and verifies its
published SHA-256 before execution, so it needs outbound HTTPS access.
The strict linked release build takes about eleven minutes on a cold local builder and can exceed Free or Trial build limits before its cache is warm. Use Hobby or higher for reliable source builds, or have CI publish a prebuilt image and point Railway at that image.
The release executable comes from the linked cargo proa build pipeline. Do
not replace it with a plain cargo build: only the linked pipeline resolves the
RSJS client-helper edges used by browser hydration.
Railway
Google Analytics
The site defaults to the GA4 web stream G-5XSLT8GCN3. Set
GOOGLE_ANALYTICS_MEASUREMENT_ID in the production service's runtime variables
to override that public measurement ID; an explicitly empty or malformed value
disables analytics. Restart or redeploy after changing it. No Google OAuth
token, service-account key, or Analytics API secret belongs in the site
configuration.
The shared document head uses proa_third_party::google::GoogleAnalytics on
marketing, docs, and ecosystem pages. The Google SDK waits until window.load
so it does not compete with the page's initial resources. The application
explicitly enables collection with AnalyticsConsent::Granted; the shared
crate otherwise requires a consent decision from its caller. The browser bootstrap
only loads Google's script on proa.so and www.proa.so, outside iframes, so
localhost, Railway previews, and embedded demos do not add production page
views. Google Signals and advertising personalization signals are disabled.
The normal GA4 configuration sends one automatic page view per document load.
See third-party integration docs for the
reusable API, consent handling, and loading tradeoffs.
To verify a deployment, open a production page and check for the
googletagmanager.com/gtag/js request and a google-analytics.com/g/collect
request with the configured measurement ID. Confirm the visit in the
property's Realtime report. Browser tracking protection may block these
requests.
Deployment
Production runs on Railway at https://proa.so. All Proa applications,
including proa-ui, have migrated from Coolify to Railway; Coolify is no
longer an active deployment dependency.
On main, GitHub Actions builds and checks the marketing image with Cloud
Build, publishes the verified image to private GHCR, and deploys its immutable
digest to Railway after CI passes. See
Marketing release images
for the release workflow and registry configuration.
For Railway builds from source, including native PR environments, the
checked-in railway.toml selects this crate's Dockerfile while keeping the
repository root as the Docker build context. The shared Cargo workspace
requires that context:
- Config file:
/railway.toml - Builder: Dockerfile
- Dockerfile:
/proa_marketing/Dockerfile - Health check path:
/healthz - Start command: inherited from the Dockerfile
CMD
Railway injects PORT, RAILWAY_GIT_COMMIT_SHA,
RAILWAY_ENVIRONMENT_NAME, and RAILWAY_PUBLIC_DOMAIN. The image provides the
remaining runtime defaults. In production, absolute URLs in robots.txt and
the sitemap use SITE_URL when explicitly set, then the deployment's
RAILWAY_PUBLIC_DOMAIN, and finally https://proa.so. Non-production Railway
environments prefer their own injected public domain so inherited production
variables cannot leak the production canonical into a PR sitemap.
Production temporarily sets SITE_URL=https://proa.so while the legacy
preview.proa.so alias remains Railway's injected primary domain. Standard PR
environments copy base variables, so the resolver deliberately prefers the
environment-specific RAILWAY_PUBLIC_DOMAIN outside the production
environment. Once Railway reports proa.so as the production public domain,
remove the explicit SITE_URL variable.
Do not set PORT: Railway allocates and routes to it. To override defaults,
set only the values you need:
HOST=0.0.0.0
RUST_LOG=info
STATIC_DIR=/app/static
SHOWCASE_PUBLIC_DIR=/app/showcase
PROA_CONSOLE_LOAD_TESTS=0
Use SITE_URL=https://<domain> on platforms that do not provide
RAILWAY_PUBLIC_DOMAIN, or for the temporary Railway production override
described above.
/readyz exposes Railway's commit SHA as x-proa-revision and returns
Cache-Control: no-store. PROA_BUILD_REVISION remains available as a
platform-neutral explicit override.
Link and inspect the service from the repository root:
railway link --project Proa --environment production --service proa
railway status --json
Production image deployment is controlled by the deploy-release job in
.github/workflows/ci.yml, which waits for the required CI gate and the
verified release image.
Enable native previews in Project Settings → Environments → PR
Environments. Keep the Railway-provided domain on the production service;
Railway uses it to provision a distinct domain for every temporary PR
environment. Do not enable Focused PR Environments yet: this service depends
on crates and build inputs across the repository, not just files under
proa_marketing/.
PROA_CONSOLE_LOAD_TESTS defaults to disabled for the public marketing console.
Set it to 1 only for local or internal deployments where /api/loadtest
should be available.
The public playground offers Proa and the deployed Hono React companions
(Node and Bun), filtered by each showcase's supported engines. Setting
PROA_COMPANIONS=0 leaves only Proa in the selector. The standalone playground
retains the other comparison engines for local benchmarking.
The eBay playground lives at /playground/ebay/ebay-srp and uses the interactive
renderer. The former /playground/ebay-rsjs/ebay-srp-rsjs URL redirects there.
The public Coolify deployment guide under src/docs/framework/ documents the
optional deployment preset for users who operate their own Coolify instance.
This repository's former Coolify operations are archived in
deploy/COOLIFY.md; legacy resource cleanup is tracked
in the decommissioning runbook.
Comments
New to MCP?
From what the protocol is to installing your first server, these answers cover most first questions.
An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.