# poordjaevin [Health: Active]

**Category:** 💻 Developer Tools  
**Repository:** https://github.com/Icaro0310/poordjaevin  
**GitHub Stars:** 6  
**Views:** 0  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/poordjaevin

## Description
Typed answers with calibrated confidence for agents. Local-first, no API key.

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `uvx` (confidence: high):

```json
"mcpServers": {
  "poordjaevin": {
    "command": "uvx",
    "args": ["poordjaevin"],
    "env": {
      "POORDJAEVIN_BACKEND": "",
      "POORDJAEVIN_ACP_MODEL": "",
      "POORDJAEVIN_ACP_TIMEOUT": "",
      "POORDJAEVIN_ACP_MAX_COST": "",
      "POORDJAEVIN_ABSTAIN": "",
      "DEVIN_CLI_PATH": "",
      "DEVIN_CREDENTIALS_PATH": ""
    }
  }
}
```

**Requires environment variables:** `POORDJAEVIN_BACKEND`, `POORDJAEVIN_ACP_MODEL`, `POORDJAEVIN_ACP_TIMEOUT`, `POORDJAEVIN_ACP_MAX_COST`, `POORDJAEVIN_ABSTAIN`, `DEVIN_CLI_PATH`, `DEVIN_CREDENTIALS_PATH` — the values above are empty placeholders; fill in real credentials before running (see the repository for what each one is for).

## Documentation

## What poordjaevin MCP server does

The poordjaevin MCP server adds structured decision-making tools to MCP clients. It accepts text and explicit answer schemas rather than asking a model to return unconstrained prose or manually formatted JSON. Results include typed values and confidence information, allowing an agent to use a decision as a gate or inspect its certainty before continuing.

Supported decision types include selecting one item from a supplied list, assigning an ordinal level such as low, medium, or high, answering a statement as true or false, and evaluating several typed questions in one model pass. The `gate` tool is intended for checks such as whether an action should be blocked before moving money or deleting data.

## How it works

The server provides five MCP tools: `gate(action)`, `judge(text, statement)`, `classify(text, options)`, `rate(text, levels)`, and `decide(text, questions)`. The first four handle individual decision patterns, while `decide` combines multiple typed questions in one request.

Two backends are documented. The default `acp` backend uses the model available through Devin's ACP interface and can rotate models automatically. It requires the Devin CLI, Node.js 18 or newer, and valid Devin credentials. The `nli` backend runs a local offline model after its initial download and does not require Devin credentials or an API key.

Confidence has different sources depending on the backend. The local path uses the project's calibrated approach, while ACP reports a temperature-adjusted self-reported probability from the model. Responses include `confidence_source`; ACP calls also log the model and cost for monitoring.

## Setup and configuration

Install the MCP package with `uv tool install "poordjaevin[mcp]"` or the equivalent `pipx` command. For offline use, install `poordjaevin[local,mcp]`, then run `POORDJAEVIN_BACKEND=nli poordjaevin serve`. The local installation downloads an approximately 400 MB model once before operating offline.

A stdio configuration can start the server with the `poordjaevin serve` command. Claude Desktop and Claude Code are documented as compatible clients. For ACP, the server finds the Devin CLI on `PATH` unless `DEVIN_CLI_PATH` is set, and reads the standard Devin credentials location unless `DEVIN_CREDENTIALS_PATH` overrides it.

Available settings include `POORDJAEVIN_BACKEND`, `POORDJAEVIN_ACP_MODEL`, `POORDJAEVIN_ACP_TIMEOUT`, `POORDJAEVIN_ACP_MAX_COST`, and `POORDJAEVIN_ABSTAIN`. The last option can make the server abstain below its calibrated threshold instead of returning an answer.

## Limitations and notes

The poordjaevin MCP server is read-only and is designed to judge or structure decisions; it does not execute the action being evaluated. ACP mode depends on the Devin CLI, Node.js, credentials, and the model service available through that interface. Its confidence is explicitly a self-report rather than an NLI log-probability measurement.

The project identifies itself as a community effort and states that it is not affiliated with Cognition AI. The package supports Python 3.10 or newer, and the repository is licensed under MIT.

_Full upstream README: https://allmcps.com/mcp/poordjaevin/readme_