MCParser

MCParser produces Universal Dependencies annotation (CoNLL-U) with a team of LLM agents. Trusted CoNLL-U files, such as a treebank’s training data, provide the examples. The agents annotate a prediction file, score every cell, and refine the least certain cells over several rounds.

raw text ──init──▶ prediction.conllu ──run──▶ scored, refined prediction.conllu ──evaluate──▶ scores
              (tokenise + baseline)      (scorer + one agent per column)

All access to the data goes through an MCP server over CoNLL-U files. The agents use it, and so can Claude Desktop or other projects (for example a parallel-corpus aligner).

Main features:

• One agent per CoNLL-U element: segmentation (FORM and multiword tokens), UPOS, XPOS, LEMMA, FEATS, syntax (HEAD and DEPREL decided together, or separately) and, optionally, DEPS, plus a scorer that judges every cell.

• A confidence for every cell, stored in the MISC column (Conf=upos:0.95,head:0.40,...), which decides what gets refined next.

• Cross-agent review: any agent can flag a probable mistake in a column it doesn’t own, and the responsible agent picks the flag up.

• Read-only trusted sources and writable predictions, enforced by the server, with a full edit history.

• A configurable model for each role (default anthropic:claude-opus-5-5) via pydantic-ai.

Guide

• Quickstart
  • Install
  • A held-out test on a UD treebank
  • Starting from raw text
  • Unannotated text
  • Exploring a treebank from Claude Desktop
• Concepts
  • Sources
  • Confidence
  • Review flags
  • History
  • Segmentation edits
  • Saving and conflicts
• The pipeline
  • Round 0: the baseline
  • Each round
  • Settings
  • Cost
• The agents
  • What an agent sees
  • What an agent may do
  • Answers
  • Joint syntax
  • Models and reasoning
  • Using the agents from Python
• Models and providers
  • Installing a provider
  • Credentials
  • Choosing models
  • Prompt caching
  • Cost estimates and --max-cost
• The MCP server
  • Running it
  • Claude Desktop
  • Tools
  • Using the corpus from another project
• Word alignment
  • Alignment sets
  • On the command line
  • Tools
  • From Python
• Evaluation
• Results
  • 100 Ancient Greek sentences, complete pipeline
  • Four full dev files
  • What a sentence costs
  • Caveats
• Troubleshooting
  • The run stops with a message about an API key
  • “STOPPED: the cost limit was reached”
  • Some tasks failed
  • Sentences have no HEAD, or UAS is much lower than expected
  • “modified by another process” (ConflictError)
  • Claude Desktop does not list the tools
  • mcparser evaluate disagrees with the official script
  • The cost estimate is missing or low
  • Reporting a bug

Reference

• CLI reference
• API reference
• Development
• Changelog