Sprout MCP

io.github.mepsopti由 GitHub 账号 mepsopti 发布v0.1.4更新于 Oct 24, 2026

Model-tiered content pipeline — cheap models seed, expensive models verify

已验证STDIO仅桌面数据与分析AI 与机器学习

概览

AI 生成的概览

一个本地 MCP 服务器，运行按模型分层的内容流水线：廉价模型起草、中档模型核查、昂贵模型验证，并记录来源与成本。

可以做什么
存储带来源、任务类型和资料来源的内容分块，并按置信度或项目列出待审核分块。将分块在 seed、watered、sprouted 阶段间推进或拒绝。按任务类型获取模型层级建议，并在运行时新增或更新路由规则。查看分块数量、置信度、token 用量和各模型预估花费。将已验证分块导出为 JSON，并生成用于批量验证的结构化审核摘要。安排、列出和取消定时任务，并跟踪失败尝试及退避建议。
适用场景
适合通过多模型流水线大量生成内容、希望把大部分工作留在廉价模型、只让昂贵模型做最终验证的场景。也适合需要逐块来源、置信度跟踪和成本报告的工作流。
运行要求
以 Python 包形式在本地运行，通过 uvx 启动，需要 Python/uv 运行环境。使用本地 SQLite 数据库，默认位于用户主目录下。可选环境变量用于设置数据库路径、自定义路由与定价的 JSON 配置文件，以及重试次数和退避。
安装前请注意
README 中的费用数字和节省比例是提供方自述，应自行核实。服务器会把内容分块和来源存入本地 SQLite 数据库，提交内容前请考虑敏感性。路由规则和定价可通过工具在运行时修改，定时任务会在之后执行，依赖前请先检查。

安装

在 SourceWeft 中

  1. 打开 控制台中的 Sprout MCP，将其添加到工作区。
  2. 为需要使用其工具的对话启用该服务。

Desktop only，通过 STDIO。 STDIO 服务会启动本地进程，因此需要 SourceWeft 桌面宿主。

其他 MCP 客户端

参照 仓库 中的启动说明。

README

Sprout MCP

Drop your Opus bill 80%. Model-tiered content pipeline for MCP — cheap models seed work, expensive models verify it.

Sprout routes tasks to the right model tier automatically. Haiku drafts, Sonnet fact-checks, Opus verifies. Every chunk tracks provenance, confidence, and cost.

Install

bash
uvx sprout-mcp

Or add to Claude Code's MCP config (~/.claude/settings.json):

json
{  "mcpServers": {    "sprout": {      "command": "uvx",      "args": ["sprout-mcp"]    }  }}

How It Works

  Haiku (seed)  →  Sonnet (watered)  →  Opus (sprouted)    Draft            Fact-check           Verify    $0.005/M         $0.015/M             $0.075/M
  1. Seed — Haiku drafts content cheaply (summarization, extraction, first passes)
  2. Water — Sonnet reviews and fact-checks the seeds
  3. Sprout — Opus deep-verifies only what passed Sonnet's review

Instead of running everything through Opus at $75/M output tokens, most work stays at Haiku's $5/M. Only the final verification — typically 10-20% of total work — touches Opus.

Tools (13)

ToolDescription
submit_chunkStore content with provenance (model, task type, sources)
get_review_queueList chunks needing review, filtered by confidence/project
mark_reviewedPromote (seed→watered→sprouted) or reject chunks
recommend_modelGet model recommendation for a task type
get_statsDashboard of chunk counts, confidence levels, token usage
export_chunksExport verified chunks as JSON
opus_testGenerate structured review summary for batch verification
schedule_taskSchedule tasks to run at a specific time or delay
list_scheduledView pending scheduled tasks
cancel_scheduledCancel a pending scheduled task
configure_routingAdd/update routing rules at runtime
get_cost_reportEstimated spend per model with real pricing
retry_on_errorTrack failed attempts with backoff guidance

Configuration

Environment Variables

VariableDefaultDescription
SPROUT_DB_PATH~/.sprout/sprout.dbSQLite database location
SPROUT_CONFIG(none)Path to JSON config file for custom routes and pricing
SPROUT_MAX_RETRIES3Max retry attempts before giving up
SPROUT_RETRY_BACKOFF2.0Exponential backoff base (seconds)

Custom Config File

Create a JSON file and point SPROUT_CONFIG at it:

json
{  "routes": {    "code_review": { "tier": "sonnet", "reason": "Code analysis needs reasoning" },    "translation": { "tier": "haiku", "reason": "Straightforward language task" }  },  "pricing": {    "custom-model": 10.00  }}

Default Routing Table

Task TypeTierWhy
biography_synthesishaikuFactual summarization
council_descriptionhaikuHistorical summarization
document_synopsishaikuContent summarization
json_validationhaikuStructural verification
summarizationhaikuGeneral summarization
data_extractionhaikuStructured extraction
fact_check_first_passsonnetCross-reference claims
code_reviewsonnetCode analysis
fact_check_finalopusDeep factual verification
theological_analysisopusDomain expertise required
complex_analysisopusDeep reasoning required

Unknown task types default to haiku — start cheap, escalate if needed.

Example Workflow

You: Use recommend_model for "biography_synthesis"Sprout: biography_synthesis → haiku-4.5 (Factual summarization)
You: Use submit_chunk to store the Haiku outputSprout: Stored chunk abc12345 [seed] for person-001.biography
You: Use get_review_queue to see what needs fact-checkingSprout: 1 chunk pending review
You: Use mark_reviewed to promote after Sonnet fact-checks itSprout: Chunk abc12345 → watered (verified by sonnet-4.6)
You: Use get_cost_reportSprout: haiku-4.5: ~1,300 tokens (1 chunk) — $0.0065        Total: $0.0065

Development

bash
git clone https://github.com/mepsopti/sprout-mcp.gitcd sprout-mcpuv sync --extra devuv run pytest

Support

If Sprout saves you money on your AI bill, consider buying me a coffee:

[Buy Me A Coffee]

License

MIT

mcp-name: io.github.mepsopti/sprout-mcp

来源：README.md，提交 65186aa

工具

0
SourceWeft 尚未在沙箱中启动过该软件包，因此其工具未知。

版本历史

1
  1. v0.1.4最新Sep 16, 2026

评分

已满足适用检查项的 50%

该评分衡量此收录条目的完整程度，依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价，也不表示服务器安全或当前可用：可连通性是带日期的检查结果，评分也不包含安装量或用户评价。

  • 近 30 天内可连通25 分

    请保持端点能够响应 MCP 握手；SourceWeft 每 30 天探测一次。

  • 已声明认证方式10 分

    请在注册中心条目中声明凭据请求头，或保持端点能够响应探测，以便判定其认证方式。

  • 已列出工具15 分

    请让服务器在无凭据时响应 tools/list，或发布可在沙箱中无凭据启动的软件包。

  • 描述必需15 分
  • 图标10 分
  • 文档15 分
  • 许可证10 分