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):

config.json

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

How It Works

Code

  Haiku (seed)  →  Sonnet (watered)  →  Opus (sprouted)
    Draft            Fact-check           Verify
    $0.005/M         $0.015/M             $0.075/M

Seed — Haiku drafts content cheaply (summarization, extraction, first passes)
Water — Sonnet reviews and fact-checks the seeds
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)

Tool	Description
`submit_chunk`	Store content with provenance (model, task type, sources)
`get_review_queue`	List chunks needing review, filtered by confidence/project
`mark_reviewed`	Promote (seed→watered→sprouted) or reject chunks
`recommend_model`	Get model recommendation for a task type
`get_stats`	Dashboard of chunk counts, confidence levels, token usage
`export_chunks`	Export verified chunks as JSON
`opus_test`	Generate structured review summary for batch verification
`schedule_task`	Schedule tasks to run at a specific time or delay
`list_scheduled`	View pending scheduled tasks
`cancel_scheduled`	Cancel a pending scheduled task
`configure_routing`	Add/update routing rules at runtime
`get_cost_report`	Estimated spend per model with real pricing
`retry_on_error`	Track failed attempts with backoff guidance

Configuration

Environment Variables

Variable	Default	Description
`SPROUT_DB_PATH`	`~/.sprout/sprout.db`	SQLite database location
`SPROUT_CONFIG`	(none)	Path to JSON config file for custom routes and pricing
`SPROUT_MAX_RETRIES`	`3`	Max retry attempts before giving up
`SPROUT_RETRY_BACKOFF`	`2.0`	Exponential backoff base (seconds)

Custom Config File

Create a JSON file and point SPROUT_CONFIG at it:

config.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 Type	Tier	Why
`biography_synthesis`	haiku	Factual summarization
`council_description`	haiku	Historical summarization
`document_synopsis`	haiku	Content summarization
`json_validation`	haiku	Structural verification
`summarization`	haiku	General summarization
`data_extraction`	haiku	Structured extraction
`fact_check_first_pass`	sonnet	Cross-reference claims
`code_review`	sonnet	Code analysis
`fact_check_final`	opus	Deep factual verification
`theological_analysis`	opus	Domain expertise required
`complex_analysis`	opus	Deep reasoning required

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

Example Workflow

Code

You: Use recommend_model for "biography_synthesis"
Sprout: biography_synthesis → haiku-4.5 (Factual summarization)

You: Use submit_chunk to store the Haiku output
Sprout: Stored chunk abc12345 [seed] for person-001.biography

You: Use get_review_queue to see what needs fact-checking
Sprout: 1 chunk pending review

You: Use mark_reviewed to promote after Sonnet fact-checks it
Sprout: Chunk abc12345 → watered (verified by sonnet-4.6)

You: Use get_cost_report
Sprout: haiku-4.5: ~1,300 tokens (1 chunk) — $0.0065
        Total: $0.0065

Development

bash

git clone https://github.com/mepsopti/sprout-mcp.git
cd sprout-mcp
uv sync --extra dev
uv run pytest

Support

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

License

MIT

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