Distribution verification#
The restored bounded remote service is live as of 2026-10-16; see the dated staging/production evidence. Existing local 0.3.0 is unchanged. Remote review package 0.3.0-remote.4 describes previously unseen entities and the actual output limits. Directory submission remains owner-only. The other ecosystem observations below are historical 2026-10-01 results unless a later date is stated; restoration does not establish fresh model compatibility.
| Ecosystem | Status | Evidence and scope |
|---|---|---|
| PyPI | PUBLISHED | Existing canonical 0.3.0; no package release for infrastructure-only changes. |
| Official MCP Registry local | PUBLISHED | Active historical 0.3.0 and current 0.3.1 metadata both retain PyPI 0.3.0 stdio. |
| Official MCP Registry remote | PUBLISHED | Current 0.3.1 is active/isLatest and adds the production Streamable HTTP URL. |
| Wikimedia Toolhub | PUBLISHED | Existing listing preserved. |
| Context7 | PUBLISHED | Existing distribution preserved. |
| Smithery Local | PUBLISHED | Existing canonical Local listing and eight tools preserved. |
| Smithery Remote | SUPPORTED_NOT_TESTED | Publisher UI access is now verified. Production URL, remote policy and setup link are published in the existing Local listing description. Simultaneously active Local and Remote release registration remains unverified; no Local replacement or duplicate listing. |
| mcpservers.org | PUBLISHED | Existing listing preserved; this does not claim its pending text update was approved. |
| OpenAI / ChatGPT / Codex | NOT SUBMITTED | Restoration package 0.3.0-remote.4 prepared for the existing plugin and individual publisher ALEKSANDR PROKOFYEV. Public endpoint/domain preserved. Live production restoration and package metadata are separate from portal scans and review. Six legal attestations remain owner-only; no submission or directory approval. |
| Claude | PENDING_REVIEW | Directory portal shows Submitted (done), In review (current), and all eight tools. Separate custom connector also discovers eight read-only tools. No public listing or model-selection test claimed. |
| Gemini API | SUPPORTED_AND_TESTED | Native Interactions on gemini-3.5-flash selected kg_status and returned a function_result (interaction status requires_action). Official MCP SDK discovered eight tools. Four bounded manual function-calling tests selected and executed search, entity, audit and ambiguous search; see the distinct path and SDK limitation below. |
| Gemini Deep Research | SUPPORTED_NOT_TESTED | Native remote MCP documented; paid inference excluded by verification budget. |
| Antigravity | SUPPORTED_AND_TESTED | Native CLI 1.2.14 connected through workspace serverUrl, discovered all eight tools, and Gemini 3.8 Flash (High) selected and executed kg_status once using the existing Starter Quota. Desktop IDE and Store publication are separate and untested. |
| Gemini/Antigravity public directory/store | SUPPORTED_NOT_TESTED | A curated Store exists; this server is not listed and no open publisher submission mechanism was verified. Prepared community request was not posted or treated as official publication. |
| Microsoft Copilot | BLOCKED_ACCOUNT | Available educational tenant excluded by owner; no agent/connection. Certification requires verified publisher/program eligibility and unresolved no-auth acceptance. |
| Cursor | SUPPORTED_AND_TESTED; BLOCKED_GITHUB | Scoped workspace trust and separately authorized server approval enabled native agent mcp list-tools: all eight tools discovered at the production URL. No model/provider call in this corrected test. Marketplace still requires a public GitHub repository; anonymous checks return 404. |
| xAI/Grok | SUPPORTED_NOT_TESTED; BLOCKED_BILLING | Native API remote MCP documented and the existing owner Console/Chat API account is accessible. No separately free MCP discovery or authorized paid model call was established; no API credential extracted, model request or consumer directory publication. |
| Perplexity | SUPPORTED_AND_TESTED; BLOCKED_BILLING | Streamable HTTP/no-auth connector discovers all eight tools. The pre-existing working connection was preserved and only the newly created duplicate was deleted with specific owner consent. Model execution excluded: plan credits are zero and only purchased credits remain. No public directory publication. |
| Mistral / Le Chat / Vibe Work | SUPPORTED_AND_TESTED | Owner explicitly authorized the displayed Vibe terms/privacy step. One no-auth custom connector discovered eight tools; the Free-plan model selected and executed kg_status once with one-time approval. All tools retain Ask permission. No public directory listing. |
| Amazon Q Developer | SUPPORTED_NOT_TESTED | Native remote HTTP/no-auth configuration documented; CLI unavailable for a test. |
OpenAI and Claude native discovery were tested, including Claude's separate custom-connector UI. Both submissions were in review at the October 1 closeout; the October 15 OpenAI readback now reports Changes required. Claude's later review outcome was not rechecked. No public listing is claimed here. The official MCP SDK 2.2 verified production protocol, schemas, status, search and bounded rejections. Closed staging exercised all eight tools, ambiguity, wrong identity, provider failure and injected daily/burst thresholds. Microsoft model tool selection remains untested.
Model and discovery closeout#
The existing Gemini Free-tier project was used without enabling billing. Native
Interactions remote MCP selected and executed kg_status on gemini-3.5-flash;
the response contained function_call and function_result, confirmed Google KG
disabled, and reported requires_action, so a completed interaction is not claimed.
gemini-3.8-flash returned a temporary high-demand HTTP 503.
The official google-genai 2.26.0 tools=[ClientSession] path failed locally:
object configuration deep-copied an asyncio Future, while dictionary configuration
reached a schema filter that called .items() on boolean additionalProperties.
The latter was reproduced offline against the public schema. The endpoint and
resolver were not modified to accommodate these SDK errors.
Four supplementary tests used Google's documented manual function-calling cycle and official MCP SDK 1.30.0 over Streamable HTTP. Tool declarations came directly from live MCP schemas; automatic execution was disabled, one selected call was validated and executed per case, and its actual result was returned to the model. This is a client test harness, not a deployed proprietary compatibility wrapper, and is not represented as four native Interactions tests:
| Example | Model-selected tool | Actual result |
|---|---|---|
| Tate Modern | kg_search | ok; two candidates |
| Q193375, selected P31 facts | kg_entity | ok |
| Existing Q193375 identity audit | kg_audit_identity | INSUFFICIENT_EVIDENCE / KIND_UNVERIFIED; no invented confirmation |
| Victoria without local evidence | kg_search | ambiguous; three candidates; model preserved uncertainty |
An initial Free-tier HTTP 429 prevented the ambiguous call before tool execution;
one later attempt succeeded. No failed call was silently treated as success.
Mistral and Antigravity CLI separately selected and executed one kg_status each; Perplexity and Cursor
are discovery-tested only. OpenAI and Claude were in review at that closeout;
the later OpenAI remediation state is recorded in the table above.
The deployment acceptance phase used all 30 permitted Wikidata HTTP attempts,
with zero Google KG and paid LLM calls. The subsequent closeout allows 20
additional Wikidata attempts exclusively for Gemini. The four tests reserved
20 attempts conservatively; responses report three each for the three search/entity
calls and two successful shared batches for audit (11 total). No further provider
tests were made. Google KG remains zero; the temporary Gemini key file was removed.
An incorrect Cursor launcher invocation unintentionally entered Agent mode
once during closeout. Its billing outcome is unknown and it is not MCP
compatibility evidence; zero model usage is not claimed for that phase.
The corrected native CLI test now discovered eight tools after narrowly scoped approval.
Deep Research and the Antigravity API agent require authorized paid inference;
neither was called. The separate Antigravity desktop IDE was unavailable, but the
owner subsequently opened CLI 1.2.14. In an isolated temporary workspace,
.agents/mcp_config.json declared only the public wikidata_knowledge server.
The native /mcp manager connected and discovered eight tools. /usage showed
100% weekly Gemini quota available on the existing Starter Quota; billing was
not enabled. Gemini 3.8 Flash (High) selected and executed kg_status once,
reported Wikidata available and public remote Google KG disabled, and listed
all eight tool names. No search, entity call, command or file edit was requested
from the model. Wikidata and Google KG attempts for this test: zero. Global MCP
configuration was not changed. CLI mcp list reports global configuration;
workspace discovery was verified in the interactive /mcp manager.
Smithery publisher UI access was subsequently verified. The existing canonical listing description now publishes the direct production remote URL, Wikidata-only policy, daily UTC admission budget and setup/privacy link. Anonymous HTML readback confirmed the new URL, Local installation and all eight tools. No release was submitted: the URL-release form and current metadata API do not establish an additive operation that preserves simultaneous Local and Remote installation. The earlier publisher-account blocker is resolved; safe transport registration remains an unverified platform capability, not a claimed platform-wide prohibition. The educational Microsoft environment remains excluded. No proprietary compatibility wrappers were built, and the production service was not redeployed.
Current setup sources: OpenAI submission, Claude directory, Gemini remote MCP, Deep Research, Antigravity MCP, Antigravity CLI, Microsoft certification, Cursor publishing, Smithery publishing, xAI, Perplexity, Mistral, Amazon Q.