Skip to content
MCP server

IRP

Lets a coding agent rebuild a paper's code and verify each claim from an evidence log

Runs locallyStatus and history
Install
Pricing
Free
Platforms
CLI, self-hosted
Runs
Locally
Version
0.6.0
Listing updated
today

Similar MCP servers

About IRP

IRP is a free MCP server for developer tools. It runs locally on your machine and is started by your AI client.

IRP does the mechanical work around paper reproduction and leaves the thinking to your agent (Claude Code, Codex or Cursor). The agent reads the paper and writes the code, while IRP fetches the arXiv source or PDF, records every claim, builds an environment, runs commands and extracts metrics. IRP itself never calls a model. A hash-chained, append-only evidence log feeds the final report, so a person can check each verdict without trusting the agent.

The report lists every claim with the paper's number and the measured number, and gives one of four verdicts: reproduced, close, not reproduced or could not run. It also includes the guesses made where the paper was vague, a map from equations to code lines and provenance such as file hashes, git commit, seeds and every command's exit code. The workflow rules forbid typing a number by hand, counting shortened runs or changing the method to get a run to pass.

It needs Python 3.11 or later on macOS or Linux, uses uv or Docker for environments, and can run on a paid GPU provider only after a cost estimate and approval. It covers computational papers, not wet-lab or weapons-related work. Every tool also works as a terminal command.

Install

Run in a terminal

claude mcp add irp -- uvx irp-agents

Add to claude_desktop_config.json

{
  "mcpServers": {
    "irp": {
      "command": "uvx",
      "args": [
        "irp-agents"
      ]
    }
  }
}
Install in CursorOpens Cursor, which asks you to confirm. Or add it by hand:

Add to ~/.cursor/mcp.json

{
  "mcpServers": {
    "irp": {
      "command": "uvx",
      "args": [
        "irp-agents"
      ]
    }
  }
}
Install in VS CodeOpens VS Code, which asks you to confirm. Or add it by hand:

Add to .vscode/mcp.json

{
  "servers": {
    "irp": {
      "type": "stdio",
      "command": "uvx",
      "args": [
        "irp-agents"
      ]
    }
  }
}

Add to ~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "irp": {
      "command": "uvx",
      "args": [
        "irp-agents"
      ]
    }
  }
}

Add to your client's MCP settings (most use this shape)

{
  "mcpServers": {
    "irp": {
      "command": "uvx",
      "args": [
        "irp-agents"
      ]
    }
  }
}

Using several servers? Build one config file with the MCP config generator.

Tools this server provides (17)

Taken from the project documentation; the running server may differ.

fetch_paper
Downloads and hashes an arXiv source or PDF and splits it into sections.
get_section
Returns one section's text with file and line numbers.
get_page_image
Returns a page as an image for equations and figures.
add_claim
Records a claim with metric, value, location and success bar.
list_claims
Lists recorded claims.
add_guess
Records a missing detail and the choice made.
map_method
Links an equation or algorithm step to a file and line.
create_env
Creates a uv environment or Docker image with a lock file.
run
Runs one command and saves its outputs and metadata.
run_status
Checks on a run.
run_remote
Runs on a paid GPU provider after cost approval.
extract_metrics
Pulls numbers from logs and results with file and line.
compare
Compares a claim with a measurement and gives a verdict.
add_note
Adds a note to the evidence.
report
Writes the report from the evidence log.
verify
Rechecks the evidence hash chain.
study_status
Shows the progress of a study.

What it does

Paper intake
Fetches arXiv LaTeX or PDF, hashes it and splits it into sections.
Claim tracking
Records each claim's metric, value, location and success bar.
Isolated environments
Builds uv environments or Docker images with lock files.
Logged runs
Saves output, exit code, time, memory, seed and commit for each run.
Evidence-based report
Writes report.md and report.json from the log and rechecks the hash chain.
CLI parity
Every MCP tool is also a terminal command.

Good for

  • Reproduce a paper's results table from the paper alone
  • Audit whether published numbers hold up
  • Document guesses where a paper omits details
  • Produce a verifiable report for a reproduction study

Pricing

Free

Runs locally; paid GPU compute only with approval and a cost estimate

IRP FAQ

Which agents does it set up?

Claude Code, Codex and Cursor via uvx irp-agents setup; other MCP clients can register it manually.

Does IRP call an LLM?

No. The README says it never calls a model.

What platforms does it need?

Python 3.11 or newer on macOS or Linux.

What papers are supported?

Computational ones: machine learning, algorithms and simulations.

Is IRP free?

Yes. IRP is free to use.

How do I add IRP to Claude Code?

Run this command in your terminal: claude mcp add irp -- uvx irp-agents

What are the best alternatives to IRP?

The closest alternatives to IRP are REA, Context7, Chrome DevTools MCP, Codebase Memory and GitHub.

Where this comes from

  • Description: written from the project's README.
  • Install steps: from the official MCP Registry entry.

How AIPoweredBot checks servers