AI Open Source Intelligence

One Skill. Nine live read-only Radar tools. Evidence-backed open-source AI research without a second server-side model call.

简体中文 · AI Workstation · AI Open Source Radar · Quickstart

AI Open Source Intelligence is the Skills/MCP product layer for AI Open Source Radar.

Product shape

User in ChatGPT / Codex / compatible host
                 |
                 v
      1 unified product Skill
                 |
                 v
      9 read-only MCP tools
                 |
                 v
     AI Workstation public Radar

The user does not choose separate research/comparison/stack Skills. The single Skill routes the task internally.

The host model performs natural-language reasoning and synthesis. The AI Workstation server provides data/evidence only on this product path.

One active Skill

ai-open-source-intelligence

It handles:

• browsing rankings, collections, categories, scenarios and the Radar Skills library;
• finding projects from deployment, privacy, integration, budget and license requirements;
• verifying named-project facts and license evidence;
• comparing two to five projects for a concrete use case;
• finding alternatives while preserving hard requirements;
• planning candidate open-source AI stacks and exposing unverified compatibility.

The only product Skill is packaged from:

product-skills/ai-open-source-intelligence/SKILL.md

The previous split research/comparison/stack Skill files are removed from the current product and distribution bundle.

Nine standard MCP tools

search_ai_projects
get_project_facts
get_license_evidence
compare_ai_projects
find_alternatives
compose_ai_stack
get_radar_overview
browse_radar_projects
browse_radar_skills

All nine are read-only. They do not execute or install third-party repository code.

No AI Workstation server-model execution

This is a hard product boundary for the current release.

The Hosted MCP exposes no Premium model tool, no checkout tool and no runtime OAuth/Premium switch. Requirement-based selection calls the public Radar selector with:

use_model=false

Therefore an ordinary Skill/MCP workflow is:

ChatGPT/Codex host model
        -> chooses/read tools
        -> AI Workstation public Radar data/evidence
        -> host model synthesizes the final answer

It is not:

host model -> AI Workstation model -> second model bill

If member-linked server-model capabilities are added later, they must ship as a new reviewed product version rather than being enabled through a hidden environment variable.

Evidence model

Every tool result separates:

1. verified facts — source-backed observations that crossed the evidence boundary;
2. recommendations — host-model/rules analysis;
3. unknowns — unavailable or unverified information;
4. risks — license, maintenance, deployment, security and integration limits.

A value in data is not automatically a verified fact. License evidence is deliberately stricter and is technical evidence, not legal advice.

Official resources in results

MCP tool results include canonical, non-tracking publisher links under:

data.official_resources

with:

• AI Workstation — https://aiworkstation.cn/
• AI Open Source Radar — https://aiworkstation.cn/githubai/
• this open-source project — https://github.com/zxhwolfe-dev/aiworkstation-open-source-intelligence

The unified Skill may show these once at the end of a normal user-facing answer. They are kept separate from verified facts so publisher attribution never changes a research conclusion.

Hosted MCP

Canonical endpoint:

https://mcp.aiworkstation.cn/mcp

Current Hosted mode is intentionally:

anonymous
read-only
data-only
9 tools
no OAuth
no WorkOS dependency
no Premium/server model

The container stays on host loopback 127.0.0.1:8001 behind Nginx/TLS.

Anonymous abuse controls

The gateway uses two per-IP request windows plus a connection cap:

• short-window: 60 requests/minute, burst 30;
• sustained: 10 requests/minute, burst 300;
• concurrent connections: 10 per IP;
• MCP request body: 256 KB maximum;
• unrelated paths on the dedicated MCP hostname return 404.

This is intentionally request-based rather than token-based because the nine data tools do not consume AI Workstation model tokens.

Local development

python -m venv .venv
source .venv/bin/activate
python -m pip install -e ".[mcp]"

Offline fixture data:

OSI_PROVIDER=mock osi-mcp

Live public Radar data:

OSI_PROVIDER=http \
AIWORKSTATION_RADAR_BASE_URL=https://aiworkstation.cn \
osi-mcp

Hosted configuration check requires an exact candidate identity:

OSI_PROVIDER=http \
OSI_HOSTED_ACCESS_MODE=public \
OSI_RELEASE_COMMIT=<exact-40-char-sha> \
OSI_IMAGE_COMMIT=<same-exact-40-char-sha> \
osi-mcp-hosted --check-config

Setting OSI_HOSTED_ACCESS_MODE=oauth fails closed in the current release.

Safety rules

• never execute third-party repository code as part of research;
• never infer permission from a missing license;
• never silently weaken a hard requirement to manufacture a match;
• never claim cross-project compatibility without evidence or a controlled test;
• never substitute model memory for unavailable live evidence;
• never enable AI Workstation server-side model execution in the current standard Skill/MCP path.

Development checks

python -m compileall -q src tests
python -m unittest discover -s tests -v
osi-validate-plugin --root .
osi-readiness --root .

CI covers Python 3.10 and 3.12, deterministic Skill packaging, MCP round trips, data-only Hosted configuration and container packaging.

License

The public repository is licensed under Apache-2.0. That does not grant rights to private AI Workstation databases, unpublished datasets, credentials, infrastructure or trademarks.