The Reexpress MCP server is designed for tool-calling LLMs (e.g., Claude) and MCP clients running on macOS (Tahoe 26 or later on Apple silicon) or Linux. It is a directly pluggable solution that adds frontier statistical verification capabilities to complex LLM pipelines as well as everyday search and question answering (for software development and data science scenarios), and is described as the first reliable, statistically robust AI second opinion for AI workflows.
Purpose and main capabilities: After installing the MCP server, add a Reexpress prompt at the end of the chat text, and the tool-calling LLM will use a pretrained Reexpress similarity-distance-magnitude (SDM) estimator to check its response. This estimator integrates gpt-5.5-2026-04-23, gemini-3.1-pro-preview, and gemini-embedding-2, and combined with the output of the tool-calling LLM, computes a robust estimate of predictive uncertainty against the training and calibration examples of the OpenVerification1 dataset. After verification is complete, users can call the ReexpressAddTrue or ReexpressAddFalse tools so that subsequent Reexpress calls dynamically incorporate the updates; when more substantial modifications are needed or another underlying LLM is desired, the training scripts included with the repository can also be used for full retraining.
Usage: In addition to providing users with instruction-based output confidence estimates, the tool-calling LLM itself can use the verification output to iteratively refine answers, determine whether additional external resources or tools are needed, or confirm whether it has reached an impasse and needs further clarification from the user, which is the so-called "reasoning with SDM verification." Data is sent to Azure/OpenAI and Google only through standard LLM API calls, among which gemini-3.1-pro-preview calls obtain standard web search access through the API; all processing of the SDM estimator is performed locally. File access uses a simple, conservative but effective mechanism: users explicitly specify the files to be sent via ReexpressDirectorySet() and ReexpressFileSet().
Version 2.5.0 implements a nested estimator, running the calibration algorithm on class and prediction conditional accuracy at decreasing probability thresholds, so that the vast majority of calibration points can be assigned to a region with an accuracy estimate greater than 0.5, and points not assigned can be regarded as out of distribution; the most conservative region retains its original interpretation and behavior. The codebase has been streamlined, the language model post-training code has been removed, and a separate repository for fine-tuning the underlying network weights will be released separately. The rest of the released model is the same as 2.4.x and was calibrated using the same data with gpt-5.5-2026-04-23 and gemini-3.1-pro-preview as generative models.
Applicable scenarios and limitations: The target audience is developers and data scientists, assuming some familiarity with LLMs, MCP, and command-line tools. The system requirement is the ability to run a small 3-million-parameter PyTorch model locally (note that this is 3 million, not 3 billion), and the computational demand is extremely low. Linux and macOS are supported. Users should add other MCP servers only from trusted sources, and should note that other MCP tools may alter this server's behavior in unexpected ways. The repository contains configuration, usage documentation, output HTML, guides, an FAQ, training and calibration data descriptions, evaluations on OpenVerification1, a system demonstration paper, and a 2.4.0 model card, among other materials. Pricing and specific charging methods are not yet provided.accuracy estimate greater than 0.5, and points not assigned can be regarded as out of distribution; the most conservative region retains its original interpretation and behavior. The codebase has been streamlined, the language model post-training code has been removed, and a separate repository for fine-tuning the underlying network weights will be released separately. The rest of the released model is the same as 2.4.x and was calibrated using the same data with gpt-5.5-2026-04-23 and gemini-3.1-pro-preview as generative models.
Applicable scenarios and limitations: The target audience is developers and data scientists, assuming some familiarity with LLMs, MCP, and command-line tools. The system requirement is the ability to run a small 3-million-parameter PyTorch model locally (note that this is 3 million, not 3 billion), and the computational demand is extremely low. Linux and macOS are supported. Users should add other MCP servers only from trusted sources, and should note that other MCP tools may alter this server's behavior in unexpected ways. The repository contains configuration, usage documentation, output HTML, guides, an FAQ, training and calibration data descriptions, evaluations on OpenVerification1, a system demonstration paper, and a 2.4.0 model card, among other materials. Pricing and specific charging methods are not yet provided.