Agent MCP — Live AILANG Knowledge for Coding Agents

mcp.ailang.sunholo.com is a remote Model Context Protocol server that exposes ~21 typed tools so AI coding agents can query AILANG documentation, stdlib, examples, design docs, benchmarks, and prompts as structured data — instead of scraping markdown.

If you're building with AILANG and your harness supports remote MCP (Claude Desktop, Cursor, Cline, Continue, Claude Code), you can add the endpoint in one click from the landing page and your agent gains live, version-pinned AILANG knowledge.

Why use it

Without the MCP	With the MCP
Agent fetches llms.txt (huge, no filter)	Agent calls stdlib_search("file") and gets just the matching functions
Agent guesses URLs from the sidebar	Agent calls docs_nav() and gets the structured tree
Agent doesn't know which prompt fits its language version	Agent passes for_version=$AILANG_VERSION and never gets stale content
Agent has no way to learn which models pass which benchmarks	Agent calls benchmarks_for_model("claude-haiku-4-5") and sees pass-rate by category

The server is public, anonymous, read-only. There's one write tool (submit_feedback) for filing AILANG bug reports / feature requests / docs gaps directly from inside an agent session — those land in our triage queue.

Connecting

Quick: one-click deeplinks

The buttons on the landing page pre-fill the MCP config for Claude Desktop, Cursor, and Cline. Click, confirm, you're done.

Manual: add to MCP config

For other harnesses (Continue, etc.) or if the deeplink fails, paste this into your MCP config file:

{
  "mcpServers": {
    "ailang-docs": {
      "url": "https://mcp.ailang.sunholo.com/mcp/",
      "transport": "streamable-http"
    }
  }
}

For Claude Desktop the file is at ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows).

Tool catalog

All tools return JSON. Version-scoped tools take an optional for_version argument; passing an empty string resolves to "latest". Responses for version-scoped tools are wrapped in a {served_for, data} envelope so callers can detect cross-version content leak.

Language reference (version-scoped)

Tool	Purpose
ailang_versions	List all versions the server can answer for
prompt_get	Fetch the canonical teaching prompt (kind: agent/devtools/compact/empty for full)
stdlib_modules	All stdlib modules with one-line summaries + export counts
stdlib_module	Full exports for one module (signatures + docstrings)
stdlib_search	Find stdlib functions by name/signature/keyword
effects_catalog	All declared effects with capability mapping + introduction version
limitations_list	Known design limitations (Y-combinator, etc.) optionally by category

Examples & patterns (version-scoped)

Tool	Purpose
examples_list	Filter examples passing under for_version, by features/status
example_get	Full source + metadata for one example by path
example_for_concept	Best-fit example for a concept query ("how do I use effects?")

Docs & navigation (version-scoped)

Tool	Purpose
docs_nav	Sidebar tree as JSON (replaces scraping the Docusaurus sidebar)
docs_search	Full-text search across the docs Markdown

Design, roadmap, changelog (unscoped — historical record)

Tool	Purpose
design_docs_list	List planned/implemented/rejected design docs, optionally by version
design_doc	Full design doc by slug
roadmap	What's shipping next, what's been rejected and why
changelog_for_version	What changed in a given release

Benchmarks (unscoped)

Tool	Purpose
benchmarks_list	List benchmark suites and runs, optionally since_version
benchmark_run	One run's full result (code, compile_ok, runtime_ok, cost, tokens)
benchmarks_for_model	Where does a given model pass/fail?
benchmarks_compare	Pairwise model comparison (a-only, b-only, both, neither)

Public feedback (the only write tool)

Tool	Purpose
submit_feedback	Anonymous bug / feature / docs / limitation report — queued for human review

Submissions land in our ailang messages list --inbox public-feedback queue. Anonymous is fine; the value is the signal, not the identity.

Versioning model

Each Cloud Run image bakes in a snapshot of the AILANG corpus pinned to the release that built it. Multi-version is handled inside the tool calls (for_version), not at the URL — there's exactly one endpoint, mcp.ailang.sunholo.com/mcp/, and the data inside it is version-scoped.

If you ask for a version that isn't in the snapshot, you get {error: "unknown_version", nearest: "..."} rather than a silent downgrade. The CLI's ailang prompt --source auto flow uses this to fall back to its embedded copy without ever surfacing wrong-version content to the user.

Caps & determinism

• Cloud Run service runs ailang serve-api --mcp-http --routes-only --caps FS,Env
• --routes-only means only @mcp_name/@route exports register as tools — no helper leakage
• --caps FS,Env is the only ambient authority (filesystem reads from the baked snapshot, env reads for snapshot dir config). Net is added in v0.15+ when submit_feedback publishes to the existing ailang-messages Pub/Sub topic
• Tool replies are deterministic for a given image SHA — replays are byte-identical

Source

The MCP tools are written in AILANG itself (mcp_tools/ in the AILANG repo) and registered automatically via the @mcp_name/@route annotations. Adding a new tool = writing one .ail file. The Go-side framework is internal/apiserver/mcp.go.

The snapshot is built by tools/build-snapshot/ which reads stdlib .ail files, baseline benchmarks, prompts, changelog, and design docs and writes JSON. The snapshot is baked into the Cloud Run image at build time so each revision is a frozen point in time.

Design doc: M-AGENT-MCP.