DECIMER MCP Server

DecimerMCPServer

mcp-name: io.github.DocMinus/decimer-mcp-server

将 DECIMER 图像转 SMILES 功能作为工具调用提供的 MCP 服务。

本项目是对 DecimerServerAPI 中现有 FastAPI 服务的轻量适配器。 它不会直接运行 DECIMER 模型。 默认情况下，适配器会发送 JSON 请求，并在需要兼容时自动回退到表单载荷。

工具

• server_health：检查 DECIMER FastAPI 服务是否可访问。
• analyze_chemical_image：向 /image2smiles/ 发送 Base64 编码的图像，并返回结构化输出。

要求

• Python 3.10+
• 正在运行的 DECIMER API 服务（默认：http://localhost:8099）

可通过以下任一版本获取：

• GitHub：https://github.com/DocMinus/DecimerServerAPI
• Dockerhub：https://hub.docker.com/r/docminus/decimer_api

安装

cd /Users/a/dev/DecimerMCPServer
uv venv
uv sync

配置

将 .env.example 中的值复制到环境中：

• DECIMER_API_BASE_URL（默认 http://localhost:8099）
• DECIMER_API_TIMEOUT_SECONDS（默认 60）
• DECIMER_MAX_IMAGE_BYTES（默认 10000000）
• DECIMER_MCP_LOG_LEVEL（默认 INFO）

运行（stdio 传输）

uv run decimer-mcp-server

或

uv run python -m decimer_mcp_server

MCP 客户端配置示例

{
  "mcpServers": {
    "decimer": {
      "command": "uv",
      "args": ["run", "python", "-m", "decimer_mcp_server"],
      "env": {
        "DECIMER_API_BASE_URL": "http://localhost:8099"
      }
    }
  }
}

输出结构

analyze_chemical_image 返回：

{
  "ok": true,
  "smiles": "CCO",
  "reason": null,
  "api_status_code": 200,
  "api_message": null,
  "classifier_score": 0.0000012,
  "classifier_threshold": 0.3,
  "classifier_decision": "structure_like"
}

当 API 分类器未返回 SMILES 时：

{
  "ok": true,
  "smiles": null,
  "reason": "not_chemical_structure",
  "api_status_code": 200,
  "api_message": "No SMILES returned by API",
  "classifier_score": 0.99999,
  "classifier_threshold": 0.3,
  "classifier_decision": "not_structure_like"
}

开发测试

uv sync --extra dev
uv run pytest

Make 目标：

make sync
make test

冒烟测试助手

针对你的 DECIMER API 运行一次健康检查和一次推理调用：

cd /Users/a/dev/DecimerMCPServer
DECIMER_API_BASE_URL=http://chitchat:8099 uv run decimer-mcp-smoke-test --image /Users/a/dev/DecimerServerAPI/example_usage/structure.png

如果将设置保存在 .env 中，请使用以下命令加载：

uv run --env-file .env decimer-mcp-smoke-test --image /Users/a/dev/DecimerServerAPI/example_usage/structure.png

或使用 make：

make smoke

如有需要，可覆盖图像路径：

make smoke SMOKE_IMAGE=/absolute/path/to/image.png

## MCP Registry publishing

Tags matching `v*` trigger `.github/workflows/publish-mcp.yml`.

Workflow steps:
- installs `mcp-publisher`
- validates `server.json`
- calls registry publish using secret `MCP_REGISTRY_TOKEN`
- publishes slug `io.github.DocMinus/decimer-mcp-server` (case sensitive; must match registry grant)

Before tagging:
1. Update `pyproject.toml` + `server.json` versions
2. Ensure `server.json` stays valid (`uv pip install jsonschema && python validate snippet from AGENTS.md`)
3. Add GitHub repo secret `MCP_REGISTRY_TOKEN` (GitHub PAT with `repo`, `workflow` scopes)

Release flow:
```bash
git tag v0.1.1
git push origin v0.1.1

请留意 Actions 选项卡。如果发布失败，请修复问题后使用 workflow dispatch 重新运行。

## Contribution
This project was built by DocMinus with AI-assisted coding support (OpenCode/Copilot-style tooling), then reviewed and tested by the author.

## AI usage policy

- AI assistance was used for scaffolding, implementation drafts, and documentation edits.
- Final technical decisions, validation runs, and acceptance were performed by the maintainer.
- Runtime behavior should be validated with local tests (`make test`) and smoke tests (`make smoke`) before release.