Hosted AILANG executor

The hosted executor is an MCP server that lets an AI assistant write, check and run AILANG code without installing anything. It is the same toolset the local ailang_bootstrap plugin gives an agent, run on our infrastructure inside a locked-down sandbox, plus the AILANG documentation tools.

https://executor.ailang.sunholo.com/mcp/connect/

Sign in with any Google account whose email address is verified. The executor keeps your Firebase user ID, the OAuth client you signed in through, and digests (not copies) of the tokens it issued you, each until it expires.

It is also the reference example of an AILANG service published as a connector: everything on this page is built with serve-api and sunholo/mcp_oauth, as described in Publish an AILANG service as an MCP connector.

Tools

Group	Tools	What they do
Code	ailang_prompt	The current AILANG syntax prompt. Read it before writing code
	ailang_check	Type-check a program (and run its contracts through Z3) without running it
	ailang_run	Run a program in the sandbox and return its output
	ailang_eval	Evaluate one expression
	ailang_builtins	Search the builtin functions
Docs	docs_search, docs_nav, stdlib_modules, stdlib_module, stdlib_search, effects_catalog, limitations_list, examples_list, example_get, example_for_concept, prompt_get, changelog_for_version, ailang_versions, install_guide, onboarding_guide, roadmap, design_docs_list, design_doc, benchmarks_list, benchmark_run, benchmarks_for_model, benchmarks_compare	The AILANG documentation, versioned. These are the tools of the open docs server https://mcp.ailang.sunholo.com/mcp/, which needs no sign-in
Account	executor_whoami	The account and OAuth client the call was made as
	submit_feedback	Report a bug or a missing feature to the AILANG team (no sign-in needed)

Every tool except submit_feedback needs sign-in. All of them are read-only: running code changes nothing outside its own temporary directory.

The sandbox

Each ailang_run, ailang_eval and ailang_check call gets a fresh directory holding one file, main.ail, which is deleted afterwards. The program runs through ailang run --policy with a fixed policy (docker/executor-policy.toml); the caller cannot change it.

Limit	Value
Capabilities	IO only: no network, no environment, no clock, no process, no file system outside the request directory
Security mode	restricted
Time	60 seconds; the supervisor then kills the whole process group
Source size	64 KB
Output	64 KB
Entry point	main
Rate	30 calls a minute per account, with bursts of up to 30. Over the limit, the call gets HTTP 429 and a Retry-After header

A program that asks for a capability it does not have gets a named policy refusal, not a crash.

What is logged. One line per call: the tool, the outcome, the duration, the size of the code, a short hash of it (so repeated retries of the same code are visible) and the first error with the request directory removed. The code itself is never logged.

Add it to your assistant

Claude (claude.ai, desktop, mobile)

Verified 2026-11-11.

1. Settings → Connectors → Add → Add custom connector, give it a name and paste the URL above.

   

2. Authentication: Sign in when needed. OAuth client: Use Claude's published identity (CIMD). There is no client ID or secret to enter.

   

3. The first tool call opens the executor's sign-in page: sign in with Google and Allow. Every tool is read-only, so Always allow under Tool permissions is safe.

   

Claude Code

claude mcp add --transport http ailang-executor https://executor.ailang.sunholo.com/mcp/connect/

Then run /mcp in a session and choose the server to sign in.

ChatGPT

ChatGPT calls these plugins. Verified 2026-11-11 on chatgpt.com: sign-in and tool calls from an ordinary chat.

1. On chatgpt.com, open Plugins → Add → Add custom MCP server. Name it, choose Server URL and paste the URL above. Authentication: OAuth: every executor tool except submit_feedback needs sign-in. Under Advanced OAuth settings, leave the client setup method on the client metadata document (CIMD); the executor does not offer dynamic registration. Tick "I understand and want to continue" and Create as a plugin.

   

2. Connect it.

   

3. The executor's own page asks you to sign in with Google and allow chatgpt.com. It shows which app is asking and where the code goes.

   

4. The plugin is listed under Personal, and an ordinary chat can use it:

   

ChatGPT identifies itself as https://chatgpt.com/oauth/client.json and returns to https://chatgpt.com/connector_platform_oauth_redirect; the executor returns iss, so there is nothing to register.

ChatGPT desktop app. Use chatgpt.com for now. The desktop app's MCPs tab (Settings → Plugins → MCPs → Add MCP server) signs in, but on 2026-11-11 the process that runs Work tasks did not send the token it got, so tool calls were refused with 401, and ordinary chats do not see servers from that tab.

Others

Gemini Enterprise. Since v0.60.1 this server accepts pre-registered clients, which Gemini Enterprise needs, in two ways. As a custom MCP data store, its tools become actions in an Agent Designer agent. As an A2A agent ("AILANG Executor", card at https://executor.ailang.sunholo.com/.well-known/agent-card.json), Gemini Enterprise hands it a whole request, and it writes, checks and runs AILANG with Gemini and answers with the result, the program and its real output. Either way, a Gemini Enterprise admin needs a client from us: ask through submit_feedback or the AILANG repository. The agent answers a limited number of requests a day, and over the limit it says so. It can also file feedback for the user. Steps and traps (including Gemini Enterprise's daily agent-creation quota): Gemini Enterprise with per-user sign-in and Gemini Enterprise A2A agent.

Microsoft Copilot. Not on this server yet: it needs dynamic registration or a pre-registered client, and has not been tried.

Without sign-in. For the documentation tools only, use https://mcp.ailang.sunholo.com/mcp/. To run code on your own machine instead, install AILANG and the ailang_bootstrap plugin, which serves the same ailang_* tools over a local stdio MCP server.

Check it yourself

ailang mcp check https://executor.ailang.sunholo.com/mcp/connect/ --target both

This checks the server name, the tool titles and hints, that no tool takes a credential, that sign-in starts with a 401 naming the protected resource, the authorization server's metadata and that the sign-in page cannot be framed.