Skip to main content
Glama
AI Workstation Open Source Intelligence MCP Server

AI Workstation Open Source Intelligence MCP Server

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 · Product page · AI Open Source Radar · Quickstart

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

AI Open Source Radar production homepage showing rankings, project cards, and the AI Open Source Intelligence product entry

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.

Related MCP server: project-surface

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:

skills/ai-open-source-intelligence/SKILL.md
skills/ai-open-source-intelligence/agents/openai.yaml

agents/openai.yaml makes the Skill's dependency on the canonical ai_open_source_intelligence Hosted MCP explicit while keeping automatic invocation enabled. This is the Skill-level dependency contract; .mcp.json is the matching Plugin-level connection contract.

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.

Requirement tools publish the actual typed constraint contract through MCP:

{"id":"web_ui","value":true,"polarity":"required"}

Formal matches are revalidated against project-detail evidence. README excerpts may directly verify narrow Docker, self-hosting, browser UI and low/no-code claims; unresolved hard requirements become near-match blockers. A License label is exposed only as observed_license_label until a direct LICENSE-family source supports a verified license fact. Response observation time and evidence-source freshness are reported separately.

Official resources in results

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

data.official_resources

with:

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.

Use it now

The published v0.3.3 Plugin packages the unified Skill and the production Hosted MCP configuration together. Codex and the ChatGPT desktop Codex host can install both from one version-pinned marketplace entry. The public ChatGPT directory listing is still pending review. Today:

  • Codex / ChatGPT desktop users can install the complete repository Plugin;

  • ChatGPT web users can register https://mcp.aiworkstation.cn/mcp as a No Authentication developer-mode app while the public listing is pending;

  • Python users can install the matching CLI/MCP package from PyPI with:

python -m pip install \
  "aiworkstation-open-source-intelligence[mcp]==0.3.3"

See the Quickstart for exact ChatGPT, Codex and Python steps and the v0.3.3 Release for signed-off assets and checksums. The immutable v0.3.0 archive remains the earlier Skills-only artifact; the complete Plugin uses the current v0.3.3 patch identity rather than replacing it.

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.

Available Tools

9 tools
browse_radar_projectsC
Read-onlyIdempotent
Inspect

Browse rankings, collections, categories, scenarios, topics or filtered projects.

Inputs: role, layer, limit, query, +14 more. Returns: object

browse_radar_skillsA
Read-onlyIdempotent
Inspect

Browse/filter the Radar Skills library or open one Skill by ID.

Inputs: kind, sort, limit, query, +7 more. Returns: object

compare_ai_projectsB
Read-onlyIdempotent
Inspect

Compare two to five projects in one explicit decision context.

Inputs: project_ids (required), locale, context, criteria, +1 more. Returns: object

compose_ai_stackC
Read-onlyIdempotent
Inspect

Compose a candidate open-source AI stack and expose unknown compatibility.

Inputs: business_goal (required), locale, request_id, constraints, +1 more. Returns: object

find_alternativesC
Read-onlyIdempotent
Inspect

Find verified candidate alternatives while preserving constraints.

Inputs: project_id (required), locale, request_id, constraints. Returns: object

get_license_evidenceC
Read-onlyIdempotent
Inspect

Get observed license evidence; the result is not legal advice.

Inputs: project_id (required), locale, request_id. Returns: object

get_project_factsA
Read-onlyIdempotent
Inspect

Get current evidence-backed public facts for one project.

Inputs: project_id (required), locale, request_id. Returns: object

get_radar_overviewB
Read-onlyIdempotent
Inspect

Discover current rankings, collections, categories, scenarios and filters.

Inputs: locale, request_id. Returns: object

search_ai_projectsC
Read-onlyIdempotent
Inspect

Find and verify open-source AI projects from explicit requirements.

Inputs: query (required), locale, request_id, constraints. Returns: object

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A3.5/5.0

    Scored across 9 tools

    Disambiguation4/5

    Most tools have distinct purposes, but search_ai_projects and find_alternatives both retrieve candidate projects, and get_radar_overview/Browse_radar_projects overlap in rankings/categories. The descriptions clarify the different intents well enough that an agent can usually disambiguate.

    Naming Consistency5/5

    All tool names follow a consistent verb_noun snake_case pattern with clear verbs like search, get, compare, find, compose, and browse. The naming style is uniform and predictable across the entire set.

    Tool Count5/5

    Nine tools is well-scoped for an OSINT-oriented AI project server: discovery, facts, licensing, comparison, alternatives, stack composition, and radar browsing are each represented without unnecessary bloat or obvious missing categories.

    Completeness5/5

    The tool surface covers the full open-source AI project intelligence workflow from searching and browsing to getting evidence-backed facts, license checks, comparisons, alternatives, and stack composition. As a read-only intelligence domain, no create/update/delete lifecycle is expected, and the available tools do not leave dead ends.

    Maintenance

    ActivityNo data
    ResponsivenessOutdated data

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      C
      quality
      B
      maintenance
      An append-only research operations framework and read-only MCP that tracks research plans, approvals, observations, claims, failures, revisions, and contributions with source-grounded evidence, providing search, evidence fetch, and audit capabilities without direct ledger writes.
      30
      91 PyPI
      MIT
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables AI agents to query a repository's verified semantic model, including capabilities, ownership, contracts, constraints, impact, health, and verification status, without re-deriving project facts. It exposes read-only tools for inspecting, reasoning about, and verifying the provenance-tracked surface.
      2
      Apache 2.0
    • F
      license
      B
      quality
      B
      maintenance
      Enables deterministic, harness-neutral research workflows by unifying research object identity, evidence receipts, claim-evidence graphs, and append-only decision logs, along with specialized scholarly skills for literature analysis, reproducibility audits, and multi-agent review.
      18
      1
      -
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables AI assistants to search and retrieve stored open-source information through read-only MCP tools, including evidence retrieval by document or snapshot ID with byte-integrity verification.
      1
      Apache 2.0