Sorftime PickFlow MCP

Model Context Protocol server + Skill for Amazon product discovery. Powered by Sorftime API.

Previously: sorftime-pickflow-skill. Consolidated into this repository. skill/ contains the legacy Claude Code Skill (SKILL.md, methodology, config). src/ contains the MCP server (16 tools, cache, pipeline, scoring).

Tools

Cache Layer (offline)

Pipeline Layer (API + cache)

Analysis Layer

Session

Install

git clone git@github.com:zhan-1002/sorftime-pickflow-mcp.git
cd sorftime-pickflow-mcp
uv sync

Configure Claude Code

In ~/.claude.json, under the project entry, add:

"mcpServers": {
  "pickflow": {
    "command": "C:/Users/asus/AppData/Roaming/Python/Python313/Scripts/uv.exe",
    "args": [
      "--directory",
      "C:/Users/asus/sorftime-pickflow-mcp",
      "run",
      "pickflow-mcp-server"
    ]
  }
}

Requires Sorftime API configured in ~/.mcp.json.

Scoring versions

• v2_semantic is the default. It treats monthly_sales_volume as a monthly estimate (sales / 30), preserves missing inputs as missing dimensions, and returns data_completeness plus score_confidence.

• v1_legacy reproduces the historical formula used for the original 97-ASIN calibration. Use it only for comparison with earlier reports.

Traffic terms are fetched through one pagination and de-duplication path for single scoring, batch scoring, reverse traffic and ASIN comparison. Partial or failed traffic responses are reported explicitly rather than treated as a true zero-keyword result.

Private evaluation

The repository does not contain evaluation ASINs, keywords or credentials. A local CSV named test_set_parsed.csv with asin,keyword columns can be evaluated without printing identifiers:

pickflow-evaluate --data-dir "D:/private/pickflow-eval" --limit 5

Use --limit 0 for the complete private set. Add --output only when an anonymous per-case diagnostic JSON is needed; standard output remains aggregate.