Kujolang.ai MCP

Official registry Developer tools

AI-native programming language and toolchain for accountable, verifiable agent-driven development.

Install

Add this to your client's MCP configuration (Claude Desktop: claude_desktop_config.json; Cursor: .cursor/mcp.json; VS Code and others use the same "mcpServers" shape).

Remote (streamable-http)
Claude Code
claude mcp add --transport http kujolang-ai-mcp 'https://mcp.kujolang.ai/mcp'
JSON configuration (Claude Desktop, Cursor, VS Code)
{
    "mcpServers": {
        "kujolang-ai-mcp": {
            "type": "http",
            "url": "https://mcp.kujolang.ai/mcp"
        }
    }
}

Documentation

From the project's README on GitHub, refreshed weekly. Links open on GitHub.

Contents
  1. Kujolang.ai MCP Server
  2. Readiness
  3. What agents can do
  4. Quick start
  5. MCP Registry publication
  6. MCP surface
  7. Repository layout
  8. Catalog integrity and synchronization
  9. Security and performance controls
  10. Verification
  11. Production deployment contract
  12. License
  13. Release freshness
  14. Telemetry

Kujolang.ai MCP Server

Version 0.2.0 is the hardened preview of this read-only catalog server. The source release and hosted catalog have independent revisions; inspect /health and the generated provenance receipts for the current catalog. Existing operational and interoperability limitations remain in the linked review documents.

kujolang-mcp is a read-only Model Context Protocol server for discovering Kujo projects, primitives, tooling, showcases, agent skills, workflows, and reviewed installation guidance. The native runtime and catalog generator are written in Kujo. The generated Cloudflare adapter includes JavaScript for hosting and aggregate telemetry.

This repository is the Kujolang.ai catalog server, not the reusable kujolang/mcp framework. It follows that framework's conventions while remaining independently deployable and product-specific.

Readiness

The repository contains the native reference server and a generated Cloudflare Worker target. The public MCP catalog interface is bounded, stateless, deterministic, and read-only. The hosted adapter separately maintains daily aggregate analytics counters. The Worker is generated from Kujo definitions and the reviewed catalog; Cloudflare executes the generated JavaScript artifact, not Kujo source directly. This is not a claim of formal enterprise certification.

The public URL is https://mcp.kujolang.ai/mcp and health is https://mcp.kujolang.ai/health once deployment verification passes. Local use is always available.

What agents can do

  • Understand Kujo and the current public catalog.
  • List and search projects, primitives, tooling, showcases, skills, and workflows.
  • Retrieve exact records with source links, install text, versions, and scope notes.
  • Receive small deterministic stack recommendations for concrete work.
  • Inspect default, grouped, and complete installer profiles without executing them.
  • Discover WebOps capabilities without unrelated search results consuming the result limit.

All seven tools are read-only, deterministic, and idempotent. There is no write tool, arbitrary URL fetcher, filesystem-path input, shell input, package installer, deployment action, or secret-reading capability.

Quick start

Run locally

Use a Kujo runtime with run, test, and the HTTP standard library:

cd /path/to/kujolang-mcp
kujo run server.kujo -- --self-check
kujo run server.kujo

The local endpoint is http://127.0.0.1:8941/mcp; /mcp/v1 is a compatibility alias and /health is the health endpoint.

For a local Streamable HTTP client:

{
  "mcpServers": {
    "kujolang": {
      "type": "http",
      "url": "http://127.0.0.1:8941/mcp"
    }
  }
}

The server advertises MCP 2026-07-28 and handshake-era 2025-06-18. Production configuration enables strict Mcp-Method and Mcp-Name binding for 2026 requests.

Quick install via GitHub MCP Registry

Kujolang.ai MCP is listed in the GitHub MCP Registry. For a quick connection to the hosted service, open the listing, select Install MCP server, and follow the instructions for your supported client.

For manual setup, add a remote MCP server in your client with:

  • Name: kujolang
  • Transport: Streamable HTTP
  • URL: https://mcp.kujolang.ai/mcp

The hosted catalog requires no client credentials or local package installation. Use the local instructions above to run the server on your own machine.

MCP Registry publication

server.json describes the hosted service for the official MCP Registry as io.github.kujolang/kujolang-mcp. It is separate from the native runtime configuration in mcp-server.json and the Kujo manifest in mcp.manifest.json.

The registry entry is independent of this source metadata. Validate and publish each reviewed release to update the registry; committing this file does not publish it. The metadata points to https://mcp.kujolang.ai/mcp.

From this directory, install the official publisher and validate the metadata:

brew install mcp-publisher
mcp-publisher validate

To publish, authenticate with a GitHub account that owns the kujolang organization, then publish the reviewed metadata:

mcp-publisher login github
mcp-publisher publish

The hosted MCP catalog requires no client credentials or local package installation. For subsequent releases, update the metadata version to match the server release and validate before publishing. Publisher credentials must stay out of Git. Publication in this registry does not guarantee inclusion at github.com/mcp; GitHub's listing request is a separate step.

MCP surface

ToolPurpose
get_kujo_overviewExplain Kujo, catalog counts, discovery flow, and the trust boundary
list_ecosystemFilter projects, skills, and workflows by kind or category
search_kujo_catalogSearch released public catalog metadata
get_catalog_itemRetrieve one exact item and its source-backed guidance
recommend_kujo_stackReturn a compact deterministic match set for a task
get_installationReturn default, core, AI, quality, showcase, operating, or complete install guidance
list_webops_capabilitiesSearch only the WebOps capability collection

Six resources expose overview, project, skill, workflow, installation, and safety data. Three prompts guide stack selection, adoption planning, and WebOps discovery.

Repository layout

server.kujo                         Thin stable launcher
src/runtime.kujo                    Configuration, modes, and HTTP routing
src/http_security.kujo              Transport guards and security headers
src/protocol.kujo                   JSON-RPC and MCP dispatch
src/registry.kujo                   Tools, resources, and prompts
src/catalog.kujo                    Catalog validation, indexing, and search
data/catalog.json                   Reviewed deterministic catalog snapshot
scripts/sync_catalog.kujo           Kujo-only catalog synchronization
scripts/generate_worker.kujo        Kujo-only Worker generation
dist/worker.js                      Reviewed generated Cloudflare Worker
wrangler.jsonc                      Free-tier Worker deployment configuration
benchmarks/search_benchmark.kujo    Kujo-only search benchmark
tests/                              Kujo test suites and snapshots

Root-level manifests, policy, configuration, license, documentation, and the thin launcher remain at the root because package managers, operators, and repository hosts discover them there. Application implementation lives under src/.

Catalog integrity and synchronization

The bundled catalog is validated at startup for schema, item shape, unique slugs, counts, and a SHA-256 content digest. Production requests never read a sibling website checkout or fetch remote URLs.

Regenerate or verify it from a reviewed Kujolang.ai checkout:

kujo run scripts/sync_catalog.kujo -- --site /path/to/kujolang.ai
kujo run scripts/sync_catalog.kujo -- --site /path/to/kujolang.ai --check

The synchronizer rejects duplicate or malformed slugs and unexpected source URL schemes. Installer profile membership is parsed from the website's public installer instead of duplicated by hand. The catalog revision covers both records and installation profiles; it detects accidental snapshot corruption but is not a substitute for signed releases or human review.

The current snapshot contains 57 projects, 135 skills, and 45 workflow records (237 records total; 44 kits plus the Publishing House Operator). There is no authoritative standalone agent catalog in the source; agent-related projects, SDKs, skills, and workflows remain discoverable under their actual source kinds rather than being presented as invented agent records.

For the autonomous weekly agent prompt that reviews source accuracy, updates GitHub, deploys through the configured hosting workflow, and verifies the live catalog, see docs/WEEKLY_REFRESH.md.

Security and performance controls

  • Search fields are length-bounded and normalized query terms are deduplicated and capped before catalog work.
  • Search text and identity fields are indexed once when the catalog loads.
  • Host is mandatory and allowlisted; browser Origin is exact-allowlisted when present.
  • JSON requests require application/json, bounded body configuration, validated envelopes, object parameters, and deterministic errors.
  • Responses include defensive no-store, no-sniff, frame, referrer, resource, and content security headers.
  • Local rate limiting uses bounded peer-IP buckets and serial handler admission to avoid counter races; health checks do not consume the quota but share overload admission.
  • Non-loopback startup fails unless trusted ingress limits are explicitly acknowledged, and public listeners refuse the application-local limiter because public quotas and overall overload must be enforced at ingress.

Kujo 1.9.0 bounds transport buffering at 8 MiB; the catalog enforces its smaller 128 KiB UTF-8 byte limit afterward. Public native ingress must enforce the smaller cap before forwarding. The Worker bounds streamed bytes before parsing. See production hardening for current limits, compatibility changes, and unverified deployment gates.

Verification

kujo run scripts/sync_catalog.kujo -- --site /path/to/kujolang.ai --check
kujo run server.kujo -- --self-check
kujo run tests/run_all.kujo
kujo run benchmarks/search_benchmark.kujo
kujo run scripts/generate_worker.kujo -- --framework /path/to/mcp
node --check dist/worker.js

Production deployment contract

Select the production configuration explicitly:

KUJOLANG_MCP_CONFIG=mcp-server.production.example.json \
  kujo run server.kujo

The trusted ingress must:

  • terminate TLS and be the only network peer allowed to reach the Kujo listener;
  • enforce body bytes, connection/read timeouts, per-client throttling, and a global overload ceiling before proxying;
  • overwrite or discard spoofable client identity headers;
  • forward only /mcp and /health with the exact public Host;
  • provide access/error metrics, alerting, capacity limits, and incident-response routing.

The public catalog intentionally requires no user credentials. If private or tenant data, remote fetching, writes, or execution are ever introduced, this authority model must be redesigned and authentication becomes mandatory.

See SECURITY.md for the complete boundary and docs/NEXT_SESSION_REVIEW.md for the next production-readiness work list.

License

MIT

Release freshness

The catalog's version and latest_release_url identify a published component release when one exists. scope_note distinguishes later default-branch work, preview support, and operator requirements. The catalog's source_version is the website version, not the Kujo runtime version; the kujo item records the current runtime. Private projects retain empty public source/install fields.

The September 9 review covers all 86 public organization repositories, the 135 released skill records, and all 44 workflows in the 0.6.0 distribution. Its source inventory is in evidence/ecosystem-refresh-2026-09-09/.

Use Kujo 1.9.0 for native catalog generation and assertion tests. Node verifies the generated JavaScript Worker and its parity with the native implementation; it is not a runtime dependency of the hosted read-only catalog. Neither server installs or executes commands returned in catalog content.

After deploying, run node tests/production_catalog_test.mjs to compare all public catalog records and installer profiles with the reviewed generated Worker. Set RECEIPT_PATH to retain a JSON receipt. This sends read-only MCP queries and never executes installation commands.

The Kujo runtime entry identifies the published 1.9.0 release across Linux x64/arm64, macOS x64/arm64 and Windows x64, including runtime npm packages. Wave C beta and Wave D alpha remain experimental; participant SDK packages remain private/unpublished. The catalog remains read-only and grants no replay authority.

The September 29 companion refresh identifies Workcell, Ability and MCP 1.2.0. Preservation, controlled application/STDIO execution and Git correlation retain their documented experimental boundaries. Dispatch publication is separate.

The September 29 SSG refresh records SSG 1.1.0, default-on experimental public WebMCP, and the independently versioned local Ability pack 1.0.0. It also synchronizes the website's reviewed Kennel install guidance. The catalog remains read-only; it does not execute SSG builds or deploy sites.

The October 7 Tribunal refresh records Tribunal 1.0.2, its local/operator-controlled hardening and exact tagged installation command. It preserves deployment-certification boundaries and the catalog’s read-only behavior.

The October 8 Agent City addition records Agent City 0.2.0, its standalone installer, local operating scope, and source links. Use get_catalog_item with slug: "agent-city" to retrieve the record. The catalog does not submit missions or execute its install command. The Howl showcase bundle provides a source-grounded offline preview.

The October 22 Commerce refresh records Commerce 0.5.1, optional PostgreSQL processing and Square payment/recovery modules, and the exact versioned npm installation command. Advanced features retain their acceptance gates. The MCP catalog provides discovery only; it does not execute payments or install packages.

Telemetry

The hosted service collects daily aggregate usage counters, without raw queries or visitor identities. See the telemetry notice for details, retention, and opt-out instructions.

Frequently asked questions

What is Kujolang.ai MCP?

AI-native programming language and toolchain for accountable, verifiable agent-driven development.

How do I install Kujolang.ai MCP?

It runs as a remote server. In Claude Code: claude mcp add --transport http kujolang-ai-mcp 'https://mcp.kujolang.ai/mcp'. For other clients, paste the JSON configuration from this page into their MCP settings.

Does Kujolang.ai MCP need an API key?

Its registry entry lists no required key or secret. Some features may still need an account with the service it connects to.

Is Kujolang.ai MCP open source?

Yes. Its code is on GitHub under the MIT license.

Who maintains Kujolang.ai MCP?

It is published by kujolang. AI.tn lists it but does not maintain or endorse it.

From the official MCP Registry. Listing is not an endorsement: review a server's code and permissions before giving it access to your data.