User repository on GitHub
Configurable decision MCP server with AI SDK, Jev, percentage scores, and bias profile routing
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Agents & MCP
Decide MCP
Local stdio MCP server that lets an agent delegate a decision: send a decision, context and 2 to 64 choices, get back a recommended choice with per-choice percentages and an explicit statement of where those percentages came from.
- Stars
- 0
- Language
- TypeScript
- License
- MIT
- Published by
- Community
- Author
- dakdevs
- Last verified
- Oct 2026
About this project
A local MCP server that lets an agent delegate a decision to a configurable model. Send a decision, context, and choices; receive a recommended choice and percentages. The agent decides what to do next. Defaults to TypeSafe AI Jev through AI Gateway, using AI SDK 7's experimentalevaluate. Supports custom decision policies, multiple bias profiles, automatic profile routing, and individual tools per profile. Requires Node.js 22+. Add this to an MCP client that supports stdio:
Quoted from the project's README. Read the full README.
Used inJev for Software & DevTools
- Languages
- TypeScript
Highlights
- `percentageSource` separates a real provider distribution from a language model's validated estimate, or reports `unavailable`.
- Named bias profiles add their policy to the global one; routing costs a second model call while explicit profile tools always make one.
- A `router.minimumProbability` threshold falls back to `default` when the proposed profile scores too low, and router errors fail loudly.
- Results come back as both MCP `structuredContent` and JSON text, so clients on either convention can read them.
Watch out
This runs locally over stdio and is not a hosted MCP endpoint, and agents without stdio support need a separate transport integration. It reads credentials from its own process environment and does not load `.env` files automatically. `percentageSource` is the field to check: the SDK's OpenAI, Anthropic and Google evaluation adapters return a choice without a distribution, so percentages come back `unavailable` rather than as estimates, and setting `mode` to `language` is what gets you estimated numbers instead. Jev's separate confidence statistic is never substituted for a selected-choice probability. Rounded native distributions may sum to 99% or 101% and those values are preserved, and tied estimates break on input order. Routing adds a second model call, and the AI SDK packages are pinned because the experimental evaluation API can change in a patch release. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
The server refuses to blur that and reports `percentageSource` on every result. `provider-distribution` means the evaluation provider's own choice probabilities multiplied by 100, with no normalization and nothing invented. `model-estimate` means a language model was asked for probability estimates, which are validated for coverage, range and a sum of one, and are explicitly not calibrated confidence. `unavailable` means the model returned a choice with no distribution and every percentage is null alongside a warning.
An extra model call. With no profiles, `decide` makes one call. With profiles and routing enabled it makes two: one to pick a profile from its description, one to evaluate with that profile's policy. The `tools` setting decides which you get, with `routed` exposing `decide` and automatic routing, `separate` exposing `decide-default` and `decide-<profile-id>` for the calling agent to pick, and explicit tools always making exactly one call. Profiles do not vote or blend scores.
Because rounded native distributions are preserved rather than corrected. The README notes they may sum to 99% or 101% and that the values are kept as they are. For estimates the rule differs in kind: they are validated rather than silently normalized, and tied estimates use input order as the tie-breaker. Also check you are not reading Jev's separate confidence statistic as a choice probability, since the README states it is not one and is never substituted.
None for the default. It runs `typesafe-ai/jev` through Vercel AI Gateway using `AI_GATEWAY_API_KEY`, with no configuration file required, and the published package can be launched with `npx` so you do not need Bun installed. Beyond that, pass `--config <path>` or set `DECIDE_CONFIG` or `DECIDE_CONFIG_JSON`, and file and inline configuration cannot be combined. The server does not load `.env` files itself, and config refers to credentials by name through `apiKeyEnv` rather than embedding secrets.
A decision string, a context that can be a string, JSON object or array up to 100,000 serialized characters, and 2 to 64 mutually exclusive choices with unique ids. The response names the selected choice, gives every choice a percentage, and reports percentageSource, model, provider, warnings and which profile ran, returned as both MCP structuredContent and plain JSON text for client compatibility.
Gateway with AI_GATEWAY_API_KEY is the default, alongside direct TypeSafe on TYPESAFE_AI_API_KEY, OpenAI, Anthropic, Google, an openai-compatible kind that needs baseURL, and custom provider modules. Models default to mode evaluation, but the SDK's OpenAI, Anthropic and Google evaluation adapters return choices with no distribution, so on those, evaluation mode reports every percentage as unavailable.
router.minimumProbability sits in [0, 1] and defaults to 0, meaning no threshold. Set a positive value and a proposed profile whose probability is too low or unavailable falls back to default; router errors fail the request rather than silently switching policy. The result reports the selected and proposed profile, the routing percentage and source, and whether a threshold fallback occurred.
More Jev resources like this
5GitHub stars
A Claude Code function hook that sends AskUserQuestion dialogs to Jev and draws a probability per option above the dialog, so you still choose while seeing how confident the model is.
Jev Agents & MCP#community#jevbest#catalog-import
Community
BeLazy167/typesafe-modTypeScript
BeLazy167/typesafe-modTypeScript48GitHub stars
Agent harness for work that outlives one chat, with resumable turns, routed state machines, and observational memory across sessions and processes.
Jev Agents & MCP#community#jevbest#catalog-import
Community
dzhng/duet-agentTypeScript
dzhng/duet-agentTypeScript37GitHub stars
Assistant framework that stores specs, memory, and specialist agents as plain markdown files in your repo for Claude Code, Codex, Opencode, and similar runtimes.
Jev Agents & MCP#community#jevbest#catalog-import
36GitHub stars
Pi extension that strips the full Agent Skills catalog from every request and loads only the skills a Jev ranking call judges relevant.
Jev Agents & MCP#community#jevbest#catalog-import
102GitHub stars
Behavioral code search that asks Jev to score repository fragments for a plain-language question and returns exact excerpts with paths and lines.
Jev Agents & MCP#community#jevbest#catalog-import
Community
nassim-arifette/jevgrepTypeScript
nassim-arifette/jevgrepTypeScript11GitHub stars
Pi Coding Agent extension that routes only at task boundaries, using Jev judgments plus declared capabilities, availability, price estimates, and failure state.
Jev Agents & MCP#community#jevbest#catalog-import
Community
win4r/pi-jev-routerTypeScript
win4r/pi-jev-routerTypeScriptFrom the Jev community
Posts from builders shipping with Jev right now.