Skip to content

VerifyDoc as an MCP server — a trust layer for AI agents

In the agentic era, AI agents extract structured data from documents and act on it. The danger is that a fluent model returns a plausible, wrong value ($42.50 → $45.20) and the agent proceeds as if it were true. VerifyDoc's MCP server gives any MCP-capable agent a trust contract: every field comes back with a calibrated confidence, a source grounding, and an accept/review decision — so the agent auto-accepts what's confident and grounded, and escalates the rest instead of hallucinating forward.

Install & run

pip install 'verifydoc[mcp]'
verifydoc-mcp            # stdio MCP server

Connect it to an agent (Claude Desktop example)

Add to your MCP client config (e.g. claude_desktop_config.json):

{
  "mcpServers": {
    "verifydoc": {
      "command": "verifydoc-mcp"
    }
  }
}

The agent now has two tools:

  • verify_extraction(document, schema, threshold?, k?, adapter?) — document is a file path or raw text; schema is a JSON Schema (leaves may declare x-scoring: exact | numeric | semantic). Returns JSON with, per field: value, confidence, grounding (page/bbox/char_span), and decision (accept/review), plus accept/review counts.
  • list_adapters() — the available extractor backends.

What the agent gets back

{
  "n_accepted": 5,
  "n_review": 1,
  "fields": [
    {"path": "total", "value": "$1,234.50", "confidence": 0.98,
     "decision": "accept",
     "grounding": {"page": 0, "bbox": [0.10, 0.24, 0.21, 0.28], "support": 1.0}},
    {"path": "tax_id", "value": "XX-999", "confidence": 0.41,
     "decision": "review", "grounding": null}
  ],
  "summary": "5 field(s) auto-accepted, 1 routed to review. Review the flagged fields against their grounding before trusting them."
}

The agent policy becomes simple and safe: act on accept fields; surface review fields (with their grounding) to a human or a verification step.

Why this matters for agents

  • No silent hallucinations enter downstream actions — ungrounded / low- confidence values are flagged, not trusted.
  • Provenance for every value — the agent (or a human) can trace a field to the exact page region before acting on it.
  • Model-agnostic — swap the adapter (OCR pipeline, VLM API, your own) without changing the agent contract.

Publishing to the MCP registry (#34)

VerifyDoc ships a validated server.json (namespace io.github.bhaskargurram-ai/verifydoc, PyPI package verifydoc, verifydoc-mcp stdio runtime). Two publish steps remain — both need an interactive GitHub login (to prove ownership of the namespace), so run them yourself from a checkout:

1. Official MCP registry — publish with the mcp-publisher CLI:

# one-time: prove you own the io.github.bhaskargurram-ai namespace via GitHub OAuth
mcp-publisher login github
# from the repo root (where server.json lives):
mcp-publisher publish

Package ownership for the PyPI entry is verified by an mcp-name: marker in the published package. Ensure the release carries it (add to the project README / package metadata, then cut a matching version):

mcp-name: io.github.bhaskargurram-ai/verifydoc

2. awesome-mcp-servers — fork punkpeye/awesome-mcp-servers, add this line under the relevant category (e.g. Data Extraction / File Systems), and open a PR:

- [bhaskargurram-ai/verifydoc](https://github.com/bhaskargurram-ai/verifydoc) 🐍 🏠 - Trust layer for document→JSON extraction: every field returns a calibrated confidence, source grounding (page/bbox), and an accept/review decision.

(🐍 Python · 🏠 local/self-hosted.) Once both land, close #34.