MCP 服务器

AI Workstation Open Source Intelligence MCP Server

Enables researching, verifying, comparing, and composing open-source AI projects with transparent evidence and uncertainty boundaries through read-only tools.

README

AI Workstation Open Source Intelligence

A distribution layer for researching, verifying, comparing and composing open-source AI projects with explicit evidence and uncertainty boundaries.

Status

M1 Alpha / pre-release. The repository contains:

three complete Skill workflows;
a validated Skills-only Codex plugin package and repo-scoped marketplace;
six read-only project-intelligence tools;
a transport-neutral Python core;
a deterministic offline mock provider;
a fail-closed HTTP provider for AI Workstation's public Radar API;
an MCP Python SDK v2 stdio server with server-wide safety instructions;
live public-contract probes, sanitized capture, validation and offline replay;
a deterministic Skills-only alpha ZIP builder with SHA-256 checksums;
a consolidated code/live/operator release-readiness report;
automated unit, plugin, MCP, workflow and packaging tests.

This is not yet a hosted public MCP service or a public-directory release. The live HTTP provider is opt-in and must pass production contract validation before an external release.

Product boundaries

Every tool result separates:

verified source facts;
analysis and recommendations;
unknown or unverified information;
risks and limitations.

The first release is read-only. It does not execute repository code, mutate GitHub projects, save collections, authenticate users or process payments.

First Skills

open-source-project-research
open-source-project-comparison
open-source-stack-planner

First tools

search_ai_projects
get_project_facts
get_license_evidence
compare_ai_projects
find_alternatives
compose_ai_stack

Repository layout

.
├── .codex-plugin/               Skills-only plugin manifest
├── .agents/plugins/             repo-scoped marketplace manifest
├── .github/workflows/           CI, live validation and alpha packaging
├── skills/                      reusable Skill workflows
├── src/aiworkstation_osi/       core, providers, MCP and validation tools
├── schemas/                     input manifest and unified result schema
├── evals/                       bilingual tool and workflow cases
├── tests/                       unit, provider, workflow and MCP tests
├── examples/                    invocation and Codex configuration examples
├── docs/                        architecture, setup and release runbooks
├── CHANGELOG.md
├── AGENTS.md
├── SECURITY.md
└── PRIVACY.md

Local setup

python -m venv .venv
source .venv/bin/activate
python -m pip install -e ".[mcp]"
python -m unittest discover -s tests -v
osi-validate-plugin --root .
osi-readiness --root .

The default readiness command verifies the code tree only. It should report code_ready=true while keeping external_alpha_ready=false until live captures, CI results, Codex testing and human review are supplied.

Install the local Skills plugin

codex plugin marketplace add zxhwolfe-dev/aiworkstation-open-source-intelligence --ref main
codex plugin marketplace list

For a local clone:

codex plugin marketplace add /ABSOLUTE/PATH/TO/aiworkstation-open-source-intelligence

Restart the ChatGPT desktop app, open the Plugins Directory, choose AI Workstation Local Plugins, and install AI Open Source Intelligence.

The plugin currently installs the three Skills only. It deliberately does not claim a bundled or registered MCP connection. Connect the local stdio server separately until a portable bundled command or hosted MCP endpoint is ready.

See docs/plugin-packaging.md.

Offline mock usage

The default provider performs no network access:

osi-m0 provider-info
osi-m0 list-tools
osi-m0 invoke search_ai_projects \
  --arguments '{"query":"self-hosted RAG with Docker and web UI"}'

The returned MOCK_DATA risk is intentional.

Public Radar HTTP provider

Enable the public AI Workstation adapter explicitly:

export OSI_PROVIDER=http
export AIWORKSTATION_RADAR_BASE_URL=https://aiworkstation.cn
export OSI_HTTP_TIMEOUT_SECONDS=30
export OSI_HYDRATE_LIMIT=5

osi-m0 provider-info
osi-m0 invoke get_project_facts \
  --arguments '{"project_id":"infiniflow/ragflow","locale":"en"}'

The adapter:

uses public read endpoints only;
requires public snapshot identity for project facts;
rejects mixed-snapshot comparisons;
requires selector evidence status to be available or disclosed partial;
keeps near matches outside formal recommendations;
converts missing and sentinel license values to explicit unknowns;
flags non-standard license labels for manual review;
rejects leaked internal selector fields and unsafe contract shapes;
never imports private akaiagents modules.

MCP server

Run a local stdio MCP server for Codex or another MCP host:

OSI_PROVIDER=mock osi-mcp

Use the public HTTP provider:

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

The server exposes exactly six read-only tools. Its instructions lead with the fact/recommendation boundary, prohibition on repository execution, license caveat and recommended tool sequence. Tests connect in memory without opening a subprocess or network port.

For Codex setup, see docs/codex-setup.md and examples/codex-config.toml.

Validate the public contract

Run both language probes:

osi-probe --base-url https://aiworkstation.cn --locale en
osi-probe --base-url https://aiworkstation.cn --locale zh

Capture, validate and replay sanitized response shapes:

osi-capture-contracts \
  --base-url https://aiworkstation.cn \
  --locale en \
  --project-id infiniflow/ragflow \
  --output-dir tmp/public-validation/contracts-en

osi-validate-contracts --directory tmp/public-validation/contracts-en
osi-replay-contracts \
  --directory tmp/public-validation/contracts-en \
  --project-id infiniflow/ragflow \
  --locale en

The capture removes query text, credentials, client IDs and internal publication fields; bounds strings and lists; and records fingerprints instead of prompts.

A manually triggered workflow performs the complete bilingual validation chain and uploads an artifact only after validation, replay and sanitization succeed:

.github/workflows/live-contract-validation.yml

See docs/live-validation-workflow.md and docs/production-validation.md.

Build the Skills-only alpha package

Create a deterministic ZIP for invited Skills-only testers:

osi-build-alpha --root . --output-dir dist/alpha
(
  cd dist/alpha
  sha256sum --check SHA256SUMS
)

The archive contains the plugin manifests, three Skills, public documentation, changelog and an embedded file manifest with individual SHA-256 hashes. It excludes Python source, tests, CI workflows, pyproject.toml and the live MCP server, so it does not imply live Radar access.

The manually triggered packaging workflow runs all release gates before uploading the archive:

.github/workflows/alpha-package.yml

See docs/alpha-tester-guide.md and docs/external-alpha-checklist.md.

Produce the final alpha-readiness report

After downloading and reviewing both live contract directories:

osi-readiness \
  --root . \
  --contracts-en /path/to/contracts-en \
  --contracts-zh /path/to/contracts-zh \
  --ci-python310-passed \
  --ci-python312-passed \
  --codex-tested \
  --artifact-reviewed \
  --live-validation-run-id 123456789 \
  --reviewer "Reviewer name" \
  --require-external-alpha \
  --output tmp/external-alpha-readiness.json

The command validates and replays the supplied captures, but CI, Codex and human review flags remain explicit operator attestations. It never claims those gates were independently observed. Public-launch readiness intentionally remains false until software licensing, public legal/support pages and hosted MCP protections are complete.

See docs/release-readiness.md.

Architecture

User / plugin or MCP host
     |
Three packaged Skills
     |
MCP Python SDK v2 stdio server
     |
Transport-neutral ToolRegistry
     |
Mock provider OR hardened AI Workstation HTTP provider
     |
Current healthy validated Radar release

zxhwolfe-dev/akaiagents remains a read-only reference and private data production system. This repository integrates through explicit public HTTP contracts rather than importing its private Python modules.

Main documentation

Remaining M1 validation

Observe successful standard CI runs on Python 3.10 and 3.12.
Run the manual live validation workflow against the intended production origin and review its English and Chinese artifacts.
Install and test the Skills plugin through the repo marketplace.
Test the stdio server from Codex using the project-scoped configuration.
Close any public-contract differences found in production responses.
Generate an external-alpha readiness report with real run and reviewer evidence.
Select a software license, public legal pages and support contact before a broad public launch.
Add hosted Streamable HTTP only after authentication, host validation, rate limiting, abuse controls and deployment policy are designed.

License

No open-source license has been granted yet. The repository is public for pre-release inspection and development; reuse rights will be defined before the first public package release.