AI Open Source Intelligence is the Skills/MCP product layer of AI Open Source Radar, aimed at research, comparison, license verification, and tech stack planning for open-source AI projects. Its core positioning is "evidence-backed," meaning it presents verified facts, recommendations, unknowns, and risks separately, avoiding treating values in the data directly as verified facts.
In terms of use, the product is used to browse leaderboards, collections, categories, scenarios, and the Radar Skills library; find projects based on deployment, privacy, integration, budget, and license requirements; verify facts and license evidence for specified projects; compare 2 to 5 projects for a specific use case; find alternatives while preserving hard requirements; and plan candidate open-source AI tech stacks while exposing unverified compatibility.
Its main capabilities consist of 1 unified Skill and 9 read-only MCP tools. Users do not need to choose research, comparison, or tech stack Skills separately; a single Skill routes tasks internally. All 9 MCP tools are read-only and will not execute or install third-party repository code. The host model is responsible for natural language reasoning and synthesis, and on this product path the AI Workstation server side only provides data and evidence, without performing server-side model execution.
In terms of usage, the product is integrated via the MCP protocol. The current Hosted mode is unauthenticated, with the container at the host loopback address 127.0.0.1:8001, fronted by Nginx/TLS. Anonymous abuse controls include: short window of 60 requests per IP per minute, burst 30; sustained window of 10 requests per IP per minute, burst 300; 10 concurrent connections per IP; maximum MCP request body of 256 KB; irrelevant paths on the dedicated MCP hostname return 404. These limits are based on request count rather than token count, because the nine data tools do not consume AI Workstation model tokens.
In terms of installation and integration, the v0.3.3 plugin packages the unified Skill together with the production Hosted MCP configuration. Codex and ChatGPT desktop Codex hosts can install both from the same version-pinned marketplace entry; ChatGPT web users can, while the public directory is under review, register https://mcp.aiworkstation.cn/mcp as an unauthenticated developer mode app; Python users can install the matching CLI/MCP package from PyPI. The public ChatGPT directory listing is still under review.
Applicable scenarios include: needing evidence-based research on open-source AI projects rather than relying on model memory; needing to compare multiple candidate projects while preserving hard requirements; needing to verify license evidence; and needing to plan candidate tech stacks while clarifying unverified compatibility.
In terms of limitations, the current version has hard product boundaries: Hosted MCP does not expose Premium model tools, checkout tools, or runtime OAuth/Premium toggles; setting OSI_HOSTED_ACCESS_MODE=oauth fails closed in the current version. License evidence is deliberately set to be stricter and constitutes technical evidence rather than legal advice. The product follows several principles: never execute third-party repository code during research; never infer licensing from the absence of a license; never quietly weaken hard requirements just to create a match; never claim cross-project compatibility without evidence or controlled testing; never substitute model memory for unavailable real-time evidence; and never enable AI Workstation server-side model execution in the current standard Skill/MCP path. The public repository uses the Apache-2.0 license, but does not grant rights to the private AI Workstation database, unreleased datasets, credentials, infrastructure, or trademarks.
