Discover the Best AI Tools

Fast Jev Compaction for Codex scores tool history with Jev and reinjects retained passages after built-in compaction. This port is an experimental proof of concept whose author advises against real-work use; it needs TypeSafe credentials and cannot replace Codex history directly.

Jevex is an experimental agent loop where Jev selects and approves tool actions while a chat model fills arguments and writes answers. It needs both model credentials; bundled tools are read-only, and arbitrary MCP tools, persistent memory and production injection defenses are outside its scope.

Jev Judgment bundles an agent skill and Python client for closed-question checks before user questions, risky actions and retries. It needs TypeSafe credentials; absent credentials or network failures let the agent continue, and confidence does not grant permission.

TypeSafe CLI is a standard-library Python command-line client that evaluates typed Choice, Score and Noul questions over supplied state. It requires a TypeSafe key and returns structured judgments, not generated chat or code.

Narrows external MCP tool discovery for Codex using Jev decisions, with bounded search fallbacks for uncertain results. Requires Vercel AI Gateway access; task context and tool descriptions are sent remotely, and applying setup changes the local Codex integration configuration.

A Pi coding harness that combines Jev probability judgments with deterministic allow, confirm, block and steering policies. Requires Pi and configured decision services; context is transmitted after redaction, and the policy layer is explicitly not a sandbox.

pi-fast-jev-compaction removes or truncates stale tool history using TypeSafe Jev while retaining ordinary conversation verbatim and leaving the session file unchanged. Branch-aware decisions fall back to Pi summary compaction on failures or insufficient reduction. It sends context externally; wrong pruning can remove evidence or images and prefix changes can increase cache costs.

pi-jev-context folds fresh repeated file reads, filters eligible command logs with Jev and recalls saved originals as verbatim searchable chunks. Judgment or persistence failures keep the original output. Only incoming results are transformed; recall can miss lexical matches, and small experiments do not establish losslessness or broad speed and token savings.

omp-jev-compaction scores tool history through TypeSafe or OpenRouter and shortens low-relevance results while preserving ordinary conversation. Sticky decisions reduce repeated prefix changes; default mode retains the record of tool calls. Original sessions remain on disk. Provider failures fall back to omp behavior, and pruning can lose useful evidence or incur cache rebuild costs.

Nerve supervises Hermes workflows with asynchronous progress, retry and completion signals and profile-based optional modules. The main orchestrator retains final stop/continue authority. Stable releases differ from development main; local Laya/OpenJev backends are development-line features with separate runtimes, while hosted Jev needs credentials and external context processing.

pi-quiet-ask adds configurable Jev rule packs to Pi for tool-risk judgments, output warnings, completion checks, repeated failures and question triage. It records decisions and offers optional interventions, with shadow defaults for guard packs. An API key is needed; failures leave normal behavior intact, so it is not a fail-closed security boundary.

Jev Agent Skill Router batches a skill catalog, shortlists candidates and returns a typed selection, no-skill or review decision. Separate need and ambiguity checks gate routing; it does not execute skills. Live use requires TypeSafe access. Published results use reused synthetic English data and do not establish production accuracy or calibrated full-catalog probabilities.

TypeSafe Skill Router is an opt-in Hermes plugin that evaluates a request and installed skill descriptions with Jev, then suggests one relevant skill before the main model runs. It needs TypeSafe credentials and sends routing context externally. Missing keys, errors or no suitable match add nothing; a suggestion does not execute or authorize the skill.

Hermes Jev Approvals replaces Hermes smart approval review with typed Jev judgments and deterministic approve, deny or escalate rules. It serves the auxiliary approval slot, not chat generation. Configured provider credentials are required; malformed judgments escalate to Hermes review.

LiteAgents is a preview Python interface for several agent frameworks, with model access through LiteLLM. Selected integrations and provider credentials are required. Install its release wheel rather than the unrelated PyPI namesake; switching frameworks does not migrate native sessions.

Evoke provides a Rust CLI and Node SDK that select installed reflex programs from natural-language requests and expose plans and permissions before actions. Current execution is local and needs Jev credentials; cross-machine signed coordination is a future vision, not an available guarantee.

JCR resolves natural-language requests through a capability tree and returns relevant command or API documentation via MCP. It does not execute those commands. The checkout-based package needs Jev credentials; decomposing compound requests can additionally call OpenAI.

Jev Guard reviews tool-call risk, returned content and skill instructions through client hooks. It requires a model backend and capabilities vary by client. Errors allow execution by default unless fail-closed mode is configured; these checks do not guarantee attack prevention.

Jevwire provides MCP judgment tools, an embeddable decision-model library and Claude hooks. Its ordinary plugin notes arrive after execution; a separate tripwire can check before execution. Jev calls require credentials, and advisory notices are not automatic security guarantees.

Command Code Jev Nudge checks stopped agent turns for unfinished work and requests continuation when useful. It uses Command Code credentials, respects waiting-for-user and no-progress signals, and stops normally when Jev fails.

This Jev MCP server offers typed classification, scoring, checks, matching and content screening with probabilities and policy bands. It requires a TypeSafe key; classification chooses one label for the supplied state, and screening is not a security or authorization boundary.

claude-jev adds Jev-based rule checks, prompt routing and verbatim context selection to Claude Code, with a Pi compaction adapter. It requires provider credentials; hooks fail open on errors, and Claude compaction depends on experimental function hooks.

Jevcore provides typed judgments through a Node library, DeepSeek Harness plugin and MCP server, with optional safety and context gates. The default offline mock returns synthetic answers; real judgments require explicitly enabled live mode and credentials, and redaction is best-effort.

jev-code exposes classification, checks, scoring and ranking through MCP, a Pi extension and a CLI for coding agents. It sends supplied evidence to configured model providers and returns typed decisions; it does not generate code or make probabilities into proof.