Project list

Browser Use’s ultrafast agent: the decision model picks both the operation and the DOM element in one request, and a small LLM is called only when text must be typed.

A phone-side conversation copilot: reads the other party in WeChat, QQ, X or Feishu, proposes replies and fills the input box — sending stays manual. It reads the screen only, hooking nothing.

Computer use on macOS: OCR the screen, classify the next action with the decision model, click. Reported at about $0.0002 a step.

A skill suite for Hermes agents covering model routing, memory, compaction, skill selection, and computer or browser use — also installs under Claude Code and Codex.

Lets the decision model pick the click while Codex thinks and verifies — reported 5–10× faster browser operations.

An agent that plays Super Mario Bros. from structured emulator state rather than pixels.

A standalone Android agent where the decision model makes every decision, with a live studio showing decision telemetry.

A WeChat intent-detection overlay for macOS: screen reading plus a local small model to judge intent and risk, then reply candidates — strictly read-only, nothing injected into WeChat.

Control a real browser by voice: intent and target are decided in about 300 ms per spoken word, then Playwright acts — often before you finish the sentence.

Browser use driven by the decision model.

A browser extension that uses the decision model to find page clutter and remove it, with reusable template rules.

A MuJoCo robot-decision workbench.

A camera-only autonomous drone in MuJoCo with the judgement step running at 2.5 Hz.

A Claude Code plugin replacing the compaction summary with decisions: every tool call and result is scored in one request, stale ones are dropped or truncated, and everything kept stays verbatim.

An agent supervisor that uses decisions to keep coding agents on task.

LLM-authored task-specific decision harnesses, with optional full-trajectory reward reflection and GEPA evolution.

A structured-decision plugin for agent harnesses: native on DeepSeek Harness, and through iPolloWork also OpenCode and Codex Harness.

A Claude Code plugin that trims long Bash output before the model ever sees it.

A decision layer for the Pi coding agent: a measured tool-call gate plus jev_ask for typed, calibrated answers.

Guardrails for Pi that steer rather than interrupt: the decision model judges irreversible and off-task tool calls, detects stuck loops, and checks unverified done claims.

A skill for writing and improving programs that call the decision model.

A staged code-review workflow with a local dashboard.

Web search with the decision model: source selection, query understanding and relevance ranking.

Routes Claude Code tasks to the cheapest capable model by asking the decision model to choose among candidates.

A PostgreSQL extension for asking your tables questions in plain language.

Per-turn routing for Codex: every turn re-decides the model, the thinking depth and the speed mode.

A local-first MCP plugin for continuous software-quality review by coding agents.

A marketing analytics platform that moved a brand-visibility classifier off an LLM and onto boolean decisions, switched by a feature flag.

A lightweight router for models, tools and subagents.

A CLI that parses a repository into a method-level call graph, then asks the decision model to flag likely defects and check project rules written in plain English.

Grep by meaning rather than pattern: every line is scored against a meaning, and meanings combine with AND / OR / NOT.

Navigates a Neo4j graph hop by hop using a classifier over candidate outgoing relationships, with beam search over answer log-probabilities.

Semantic ifs from open models, on a single 3090 at home. Independent; not affiliated with the vendor.

A 0.6B replica: parallel decisions over dynamic candidates, no output-token decoding, with the end-to-end training pipeline.

Train a small model that chooses among a changing list of text options, one probability per option in a single pass — with Doom, chess and Wikispeedia demos.

A 395M non-autoregressive model answering typed questions with calibrated probabilities in under 15 ms, positioned as a local drop-in.

Turn any open model into a classifier, or into a compatible endpoint.

Turns any LLM into a decision model: typed decisions and calibrated probabilities are read from next-token prefill distributions, with no training.

An open, vendor-compatible decision server built on DiffusionGemma.

One-pass typed decisions with calibrated probabilities, fine-tuned from a 2B model.

Adapts local language models into compatible structured decision engines, with all three primitives produced by prefill-only binary inference.

A 0.6B replica aimed at agents: feed it unstructured state (diffs, traces, logs) and structured questions, get calibrated distributions back in one ~50 ms pass with no output tokens.

A calibrated 151M non-autoregressive decision engine that its author reports as beating both the vendor model and Laya on a public benchmark: 77.10% accuracy, 0.0636 Brier, 0.0144 ECE.

An educational inference experiment on Apple Silicon: shared context, direct candidate scoring, and local visual demos.

A self-hosted drop-in replacement.

A small open decision model: state plus typed questions, calibrated probabilities back.

Vercel’s open agent framework, which ships the decision model as the default evaluation model in its experimental evaluate path.

The official agent skill for Claude Code, Codex and compatible agents: the primitives, the patterns, and how to structure evaluations.

A Vercel Labs terminal CLI that can run the decision model as the evaluation model for its evaluate command.

A proof-of-concept MCP server for connecting the decision model to MCP clients.

A single-binary Go MCP server exposing typed judgements to Claude Desktop, Claude Code and Codex.

The official drop-in client replacement backed by OpenAI, Anthropic or compatible LLM APIs — made for comparing the decision model against chat models in the same evaluation.

The official TypeScript/JavaScript client, with inferred answer types.

The official sync and async Python client.

One trading decision every Monad block.

An input that becomes what you mean: one text box that morphs into the right UI as you type. Works offline.

An LLM plus decision-model workflow. (The public description goes no further than this.)

An active-learning CLI that uses human labels and GEPA to improve how decisions are defined.

Describe your startup idea and it decides: kill it, fix it, or ship it.

A live decision-driven trader on Hyperliquid.