# @adxtya21/decisionkit [Health: Active]

**Category:** 💻 Developer Tools  
**Repository:** https://github.com/adityajha2005/decisionkit  
**GitHub Stars:** 8  
**npm Downloads (last month):** 1177  
**Views:** 0  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/adxtya21-decisionkit

## Description
Fast, cheap, probability-scored decisions for coding agents, scripts and CI: a CLI and MCP server for the OpenAI Decisions API and Jev (TypeSafe AI)

## Tools
Capabilities this server exposes over MCP:

- **decide** — Ask an external decision API (Jev or OpenAI Decisions; the result names which) closed-form questions about a text context: binary (probability true), choice (one of listed options), score (ordered rubric, 0-based), rank (order listed candidates; derived). Returns provider probability estimates, not verified calibrated, and no rationale. Use only for judgments over text with a closed answer set. Put all related questions about the same context in one call (one round trip). NEVER put secrets, credentials, personal data, or code the user hasn't cleared into context: it is sent to a third party. If the result is an error, report it; never substitute your own answer.
- **status** — Show decisionkit's active provider and model, which DECISIONKIT_* settings it can see (names only, never values), a free connectivity check, and local usage totals (calls, errors, latency, tokens, estimated cost).

## Claude Desktop Quick Installation
Install path inferred — verify against the README before running it. Uses `npx` (confidence: medium):

```json
"mcpServers": {
  "decisionkit": {
    "command": "npx",
    "args": ["-y","@adxtya21/decisionkit"],
    "env": {
      "DECISIONKIT_PROVIDER": "",
      "DECISIONKIT_OPENAI_API_KEY": ""
    }
  }
}
```

**Requires environment variables:** `DECISIONKIT_PROVIDER`, `DECISIONKIT_OPENAI_API_KEY` — the values above are empty placeholders; fill in real credentials before running (see the repository for what each one is for).

## Documentation

## What @adxtya21/decisionkit does

@adxtya21/decisionkit turns questions about text into structured decisions. It supports four decision forms:

- Binary questions return a probability for true.
- Choice questions select from 2–255 options and return a probability for each option.
- Score questions place an item on 2–10 ordered levels and return the expected level, most likely level, and per-level probabilities.
- Rank questions order 2–20 candidates using one score per candidate.

The package is available as an MCP server, a command-line program, and a Claude Code plugin. MCP clients can use it for issue triage, API-change checks, incident severity assessment, and ranking files or other candidates against a piece of context.

## How it works

Each request sends its context and question to the single provider selected in configuration. Supported providers are Claude through the existing Claude Code login, the OpenAI Decisions API, Jev from TypeSafe AI, and Jev through OpenRouter. Results identify the provider and model; the CLI reports that information on stderr.

The MCP interface exposes a `decide` operation for decisions and a `status` operation for provider connectivity and usage information. The CLI reads context from standard input or a `--context` option. Running `decisionkit` without arguments starts the MCP server.

Decision probabilities are estimates from the selected provider and are not presented as calibrated probabilities. Ranking is derived from separate candidate scores, so nearby scores should be treated as ties. The system returns provider-specific fields under `native` and does not generate rationales.

## Setup and configuration

Node.js 20 or newer must be available on the `PATH`. Install the package globally with:

```bash
npm i -g @adxtya21/decisionkit
```

Set `DECISIONKIT_PROVIDER` to `claude`, `openai`, `jev`, or `openrouter`. The OpenAI quickstart uses `DECISIONKIT_OPENAI_API_KEY`; Claude uses the existing Claude Code login and does not require an additional API key. Provider credentials are selected explicitly rather than taken from generic `OPENAI_API_KEY` or `OPENROUTER_API_KEY` variables.

The Claude Code plugin can be installed from the project's marketplace and configured with `/plugin configure decisionkit@decisionkit`. Plugin settings include the provider, provider key, model, timeout, and local usage logging. Keys entered through the plugin are masked and stored in the operating system's secure store.

## Tools and capabilities

The CLI includes `binary`, `choice`, `score`, `rank`, and `status` commands. Binary, choice, and score operations support thresholds for automation. Exit code 0 indicates a true or sufficiently confident result, 1 indicates false or below threshold, and 2 indicates configuration, network, or provider failure.

Terminal output includes probability bars and a provider footer, while piped output uses a plain format. `--json` returns the complete result, `--pretty` and `--plain` select display modes, and `--quiet` hides the provider disclosure line. A local usage log records identifiers, timings, token counts, and estimated cost without prompt content.

## Limitations and notes

Context is sent to the configured third-party provider, so secrets, credentials, personal data, and uncleared code should be excluded. Agents call the tool only when they choose to; content is not sent elsewhere automatically. Claude Code displays the exact payload in a permission prompt, while other clients may auto-approve tools.

A standalone CLI process creates a fresh provider connection on every run, making repeated calls slower than a persistent MCP process. Failures, refusals, and invalid provider responses are errors rather than decision results. Thresholds should be selected using labeled examples appropriate to the task.

_Full upstream README: https://allmcps.com/mcp/adxtya21-decisionkit/readme_