Meta-Code

Use natural language, MCP, or Skill to find the paper's own code, method chain, and evidence boundary in a read-only knowledge package.

It does not execute candidate repository code, does not go online, and does not modify production. By default it only returns the paper's own code routes (T0 / T1, plus recall-layer URL-bearing T1). Third-party tool ecosystems and entries that "claim to have code but have no address" are hidden by default.

Installation requires two separate things:

1. Software: metacode-sci==0.1.0rc4

2. Data: GitHub Release asset metacode-data-0.1.0rc4.zip

At runtime it stays read-only and offline. The software does not download data automatically.

Installation (GitHub Release)

Requires Python 3.11+. Download from v0.1.0rc4:

• metacode_sci-0.1.0rc4-py3-none-any.whl

• metacode-data-0.1.0rc4.zip

• SHA256SUMS.txt

# 1. 核对校验和（对照 SHA256SUMS.txt）
Get-FileHash .\metacode-data-0.1.0rc4.zip -Algorithm SHA256
Get-FileHash .\metacode_sci-0.1.0rc4-py3-none-any.whl -Algorithm SHA256

# 2. 解压数据。METACODE_DATA_ROOT 必须指向解压后的 bundle 根目录
#    （该目录含 BUNDLE_MANIFEST.json 和 content_packet_v1_1_1）
Expand-Archive .\metacode-data-0.1.0rc4.zip -DestinationPath .
$env:METACODE_DATA_ROOT = (Resolve-Path .\metacode-data-0.1.0rc4).Path

# 3. 安装软件
python -m pip install .\metacode_sci-0.1.0rc4-py3-none-any.whl

# 4. 体检并启动 MCP
metacode-doctor
metacode-mcp

The distribution name on PyPI is metacode-sci. metacode is already taken by an unrelated project; do not install or upload under that name. If Trusted Publishing is enabled:

python -m pip install "metacode-sci==0.1.0rc4"

Do not cd into the data package and then run python -m; the data directory is not code.

Related MCP server: uni-kb

MCP Configuration

Official MCP Python SDK v2 stdio service, entry point metacode-mcp:

{
  "mcpServers": {
    "metacode": {
      "command": "metacode-mcp",
      "env": {
        "METACODE_DATA_ROOT": "/absolute/path/to/metacode-data-0.1.0rc4"
      }
    }
  }
}

Five read-only tools: knowledge_status, search_analysis_routes, open_research_case, trace_evidence, find_code.

Natural language routing is in skills/meta-code/SKILL.md. Codex / Claude Code integration is in docs/INTEGRATION_CODEX.md and docs/INTEGRATION_CLAUDE_CODE.md.

Three Example Questions

1. Where is the code for the ovrlpy paper itself? — should return the paper's own repository, not third-party dependencies.

2. POSST macrophage — recall-layer URL-bearing T1, navigation_depth=RELATION_ONLY, has a code address but is not yet a deep Figure/Methods mapping.

3. How many code repositories does the project have in total? — will not report a fake total repository count; it will separately give Phase2E 55, AP209, T0+T1 182, recall 170, evidence lower bound 352, heterogeneous 379.

Evidence Hierarchy

require_code=true or find_code excludes claim-only entries without a URL. 352 is the paper-own-code evidence lower bound, not already static-gated / reproducible / mature T0+T1. 379 is 209 old ARTICLE_PRIMARY (T0/T1/T2/T3 mixed) + 170 recall-confirmed, not a single homogeneous set. This corpus is not a complete top-journal library, nor a field consensus.

License

• Software: Apache-2.0

• Project-generated annotations and compilation rights: CC-BY-4.0 (only the parts Meta-Code is entitled to license)

• Third-party papers, figures, full texts, repository source code: not relicensed

Privacy and Security

• Read-only

• Offline by default

• Candidate repository code execution = 0

• No writes to production

• Public data package contains no PDFs, clones, credentials, or local absolute paths

Current version 0.1.0rc4. Distribution name is metacode-sci.