Remote MCP#

Restoration release, 2026-10-16. Bounded requests for previously unseen entities are supported. The seven exact request/result catalogue and recursive adult-category closure have been removed. Identity uncertainty, contradictions, shared quotas and Wikimedia coordination remain enforced. See current controls and measured verification.

Production endpoint: https://wikidata-knowledge-mcp.revanalex.workers.dev/mcp.

Verified on 2026-10-01 with official MCP SDK 2.2: server discovery, all eight tool schemas and status. The production search returned Wikidata Q1. All eight tools, ambiguity, wrong identity and provider failure were exercised on the closed staging deployment of the shared code. Native Claude directory connection also discovered the eight read-only tools. OpenAI's native scan found all eight tools with no issues and confirmed the endpoint domain. Client examples below describe native support; model tool selection and directory approval are separate checks. Set REMOTE_MCP_URL to the endpoint above.

The optional remote service provides zero-install, read-only access to the same canonical Python resolver as PyPI 0.3.0. Remote V1 is Wikidata-only. Google KG is disabled, including fallback and crosschecks; requesting it returns UNSUPPORTED_REMOTE_PROVIDER. Local PyPI/stdio retains optional Google KG with your own key. No user sign-in, provider key or OAuth is required.

Transport and availability#

Send one JSON-RPC message to POST /mcp using Streamable HTTP. Responses are JSON; there is no legacy SSE transport or session persistence. GET/DELETE return 405. The SDK supports legacy initialize and the current stateless protocol. See the current MCP transport specification.

All gateway HTTP admissions share 100,000 requests per UTC day. Setup, tools/list, notifications, invalid messages, unknown paths and burst rejections count before resolver execution. A SQLite Durable Object atomically increments the authoritative count. It resets on the first admitted request after UTC midnight. Concurrency and restarts do not create separate counters.

Exhaustion returns HTTP 429, JSON-RPC error PUBLIC_DAILY_QUOTA_EXHAUSTED, error.data.reset_at, Retry-After and X-Quota-Reset. The service intentionally remains unavailable until midnight UTC. Accounting failure returns 503 and never executes the resolver. There is no overflow into unlimited resolver usage. Cloudflare bills rejected ingress on Workers Paid; this admission ceiling is not a total platform billing cap.

Burst controls add a 600/minute global ceiling and an approximate 300/minute network gate. Shared networks may encounter burst rejection. Availability is best effort. Read privacy and service limitations.

Tool limits#

These schema ceilings apply to new names and QIDs as well as the examples. Selected-field projection and direct content admission still apply; there is no per-entity approval catalogue.

Each HTTP request permits at most eight upstream attempts across all phases, zero retries, 5 seconds per upstream request and 20 seconds per tool request. Input: 16 KiB, depth 12, strings/URLs 512 characters, aliases 8. Tool text: 6000 UTF-8 bytes; full response: 64 KiB. Runtime ceilings: Python core 1250 ms CPU/8 subrequests; gateway and quota object 20 ms CPU/4 subrequests. These follow Cloudflare acceptance and native client-discovery measurements; the original930.018ms sample is historical. Current measurements and caveats are in the release evidence. Raw SPARQL, unrestricted graphs and corpus batches remain local CLI workflows.

Provider concordance is not local identity proof. sameAs candidates are for review, not automatically approved public links. Preserve HOLD and AMBIGUOUS.

OpenAI, ChatGPT and Codex#

For OpenAI Responses API remote MCP, use a native tool configuration:

{"type":"mcp","server_label":"wikidata_knowledge","server_url":"REMOTE_MCP_URL","require_approval":"always"}

A portable plugin ZIP uses mcp.json:

{"mcpServers":{"wikidata_knowledge":{"type":"streamable-http","url":"REMOTE_MCP_URL"}}}

ChatGPT custom connectors require access to the appropriate account feature. Public plugin distribution additionally requires review and publisher/domain verification under OpenAI's submission process. Configuration does not mean directory publication. Restoration package 0.3.0-remote.4 replaces the earlier exact-profile metadata. The OpenAI draft remains Not submitted; scan and upload status is recorded in distribution verification. Directory approval is a separate decision. The Claude directory submission was In review on 2026-10-01; this is a separate historical observation, not a fresh approval or public listing check.

Claude#

Use Settings → Connectors → Add custom connector and enter the production URL, with no OAuth configuration. See Claude's custom connector instructions. Public distribution uses the separate Single MCP connector route in the directory submission portal. A custom connection does not approve the directory listing.

Gemini API and Deep Research#

Interactions remote MCP uses Streamable HTTP and a snake_case server name:

from google import genai

client = genai.Client()
tools = [{"type": "mcp_server", "name": "wikidata_knowledge",
          "url": "REMOTE_MCP_URL",
          "allowed_tools": [{"tools": ["kg_status"]}]}]
result = client.interactions.create(
    model="gemini-3.5-flash", tools=tools, store=False,
    input="Call kg_status once and report the remote provider configuration.")

For Deep Research, use the same tools with agent="deep-research-preview-04-2026" and background=True instead of model. Deep Research inference is billed under Google's pricing. Native Interactions selected kg_status and returned a function result on gemini-3.5-flash; its interaction status was requires_action, so a completed interaction is not claimed. Deep Research was documentation-checked only. Four additional bounded examples used Google's documented manual function-calling cycle with the official MCP SDK. The automatic tools=[ClientSession] path in google-genai 2.26.0 encountered local SDK errors; see the detailed verification receipt before choosing that path.

Antigravity#

The current custom configuration uses serverUrl in global or workspace mcp_config.json:

{"mcpServers":{"wikidata_knowledge":{"serverUrl":"REMOTE_MCP_URL"}}}

Native Antigravity CLI 1.2.14 was tested with this configuration in the workspace file .agents/mcp_config.json. The interactive /mcp manager discovered all eight tools. Gemini 3.8 Flash (High) selected and executed one kg_status using the existing Starter Quota; no provider requests or billing changes were made. CLI agy mcp list shows global configuration, so use /mcp to verify this workspace configuration. This test does not establish desktop IDE compatibility.

An official curated MCP Store exists. No open publisher submission path has been verified for this project. A working custom config is not Store approval.

Gemini Code Assist#

The current agent mode documentation still supports MCP. For VS Code, use native Streamable HTTP in ~/.gemini/settings.json:

{"mcpServers":{"wikidata_knowledge":{"httpUrl":"REMOTE_MCP_URL"}}}

Reload the IDE and use /mcp to inspect discovery. This configuration is documentation-checked; Code Assist and model tool selection were not tested.

Microsoft Copilot Studio#

Use the MCP onboarding wizard with the HTTPS endpoint, Streamable transport and no authentication. Generic consumer Copilot access is not established by a Copilot Studio configuration. Certification is a separate Partner Center workflow requiring publisher/account eligibility. No connection was created in the available educational tenant; it was excluded from verification. No Copilot Studio compatibility or certification is claimed.

Cursor#

Cursor remote MCP accepts:

{"mcpServers":{"wikidata_knowledge":{"url":"REMOTE_MCP_URL"}}}

Add this to the appropriate project/global MCP configuration. Existing local stdio and Smithery Local remain available. Remote configuration does not waive any repository requirement for the Cursor Marketplace. Native agent mcp list-tools wikidata_knowledge discovered all eight tools after scoped workspace trust and separately approved server enablement. This test used no model or provider calls. Marketplace publication remains separately blocked by its public GitHub repository requirement.

xAI and Perplexity#

xAI's API supports native remote MCP with server_url, server_label and allowed_tools. No xAI model call or consumer Grok directory publication was performed.

Perplexity custom remote connectors offer Streamable HTTP and None authentication. Where the account feature is available, use Account settings → Connectors → Custom connector → Remote, the production URL and those options. The platform asks the account owner to acknowledge custom connector risks. An authenticated form was prepared with Streamable HTTP and None. Its mandatory risk acknowledgement was accepted with explicit owner consent. Native discovery verified all eight tools. The pre-existing connector was preserved; only a newly created duplicate was deleted with specific approval. Model testing stopped because plan credits were zero and only purchased credits remained; no credits were spent for this test.

Mistral / Le Chat / Vibe Work#

Use custom MCP connectors, the production URL and no authentication. The existing Free-plan account added one connector and discovered all eight tools. With Ask permission and one-time approval, its model selected and executed kg_status once. This verifies custom connection and bounded model execution, not public directory publication.

Distribution verification status records publication, native discovery and remaining account/consent gates separately.

Recorded remote verification#

This is the historical failed-run recording. See the current safety revision and labelled production/staging replay for subsequent verification.

This replay uses actual public-endpoint SDK responses captured on 2026-10-15. Presentation timing is edited; tool latencies are measured. Discovery, resolution, selected facts, compact lookup, status and explicit Google rejection are shown. Audit and discovery returned partial provider errors and are labelled as failed positive cases. Positive Wikidata search and related cases were not rerun within this session budget. The conservative Wikidata attempt bound is 24; Google and paid LLM calls are zero. This is not a ChatGPT model-selection recording or a complete successful review case rerun. No moderation approval or public directory publication is claimed.

Download the terminal recording or read the transcript.