Jevbridge
Zero-dependency Node library with MCP and ACP adapters adding Jev decisions and gated computer use beside any LLM agent.
Overview
Jevbridge is an adapter that sits alongside the LLM you already run and gives it Jev as a decision sidecar. It evaluates one state against mixed Noul, Choice and Score questions and returns answers software can branch on, while the LLM writes plans and explanations. Backends are swappable without changing question shapes: native Jev when TYPESAFE_API_KEY is set, any LLM acting as System One (Codex, Claude, Grok, OpenCode or an OpenAI-compatible endpoint) when it is not, an offline heuristic backend for tests and CI, or automatic selection.
Confidence gates map a peaked distribution to execute, a middling one to confirm, and others to escalate or abort, so destructive clicks are not unsupervised. It speaks MCP with jev_decide, jev_gate and jev_computer_use tools, and the Agent Client Protocol over stdio so editors such as Zed and JetBrains can use Jev as a decision sidecar. Recipes cover support routing, destructive command gates, context keep or drop and GUI next action. It is a zero-dependency Node 22 library, planned for npm as @tactico/jevbridge, and MIT licensed.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Does this message ask for a refund? (support-route recipe) | — |
| Choice | Should this tool call or click execute, confirm, escalate or abort? | execute · confirm · escalate · abort |
| Choice | Which GUI element is the next action?(inferred) | — |
How it works
- Host agent calls Jevbridge via MCP, ACP or the library
- Backend (Jev, LLM-as-System-One or heuristic) answers typed questions
- Confidence gate decides execute, confirm, escalate or abort
- Agent proceeds accordingly
Strengths
- Works with or without a TypeSafe key
- Same question shapes across backends
- ACP support for editor hosts
- Zero runtime dependencies
Caveats
- LLM-as-System-One fallback lacks native calibration
- npm package pending release at time of README
Getting started
git clone https://github.com/tacticocc/Jevbridge.git && cd Jevbridge && npm install, then `node --experimental-strip-types src/cli.ts eval support-route`.At a glance
- Type
- MCP server
- Status
- beta
- Language
- TypeScript
- Runtime
- Node library / MCP and ACP servers
- Has code
- yes
- Has tests
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Agent bridges: ACP and MCP adapter that exposes Jev typed decisions to Codex, Claude, Grok, and other LLMs.
yibie/awesome-jevUnofficial ACP/MCP adapter: typed Jev decisions and computer use beside Codex, Claude, Grok, and OpenCode
AnotiaWang/awesome-decision-modelsACP and MCP adapter that bridges TypeSafe Jev with any LLM — computer use and…
wh000wh000/awesome-jev-liveACP and MCP adapter that bridges TypeSafe Jev with any LLM — computer use and typed decisions alongside Codex, Claude, Grok, and OpenCode.
hellogumbo/awesome-jevACP and MCP adapter that pairs Jev with any LLM agent, including Codex, Claude, Grok, and OpenCode, for typed decisions and computer use.
MrJev/awesome-jev
Frequently asked questions
Do I need a TypeSafe key?
No; without one it can use an LLM as a System One backend or an offline heuristic.
What is ACP support for?
So Agent Client Protocol hosts like Zed and JetBrains can use Jev as a decision sidecar.
How are risky actions handled?
Confidence gates return execute, confirm, escalate or abort.
Related projects
Accessibility-first macOS computer-use MCP server and CLI, with optional Jev guards on targets and inputs before irreversible actions.
PluginbetaPythonMCP server giving any agent Jev's classify, score, yes/no check and allow/confirm/block gating of risky tool calls.
MCP serverbetaJavaScriptCodex computer-use skill where Jev picks the next on-screen element, action, completion and risk from text candidates.
Agent skillprototypeJavaScriptClaude Code, Codex and pi plugin that hands non-writing agent steps to Jev and escalates writing or unsure steps to the LLM.
PluginbetaJavaScriptMCP server that lets agents observe and control Android phones over ADB, with Jev picking controls and judging goal completion.
MCP serverbetaTypeScriptClaude Code mod that adds $.jev so other mods can ask typed Choice, Score and Noul questions, for example to deny destructive commands.
ProjectbetaTypeScript