JarvisCore
Python multi-agent runtime with native Jev support: agents ask typed questions via a decision client separate from the text model.
Overview
JarvisCore is a production-oriented Python runtime for multi-agent systems. Agents form a peer-to-peer SWIM gossip mesh over ZMQ with no central orchestrator, claim workflow steps atomically from Redis so state survives kill -9, and use typed integration atoms with credentials injected at call time by a built-in zero-trust broker. It includes per-step traces, AutoAgents that write sandboxed code, and an epistemic ledger.
TypeSafe Jev is supported for bounded Choice, Score and Noul judgments via an optional extra. With TYPESAFE_API_KEY set, the mesh injects one shared async client as self.decisions, and AutoAgents get an evaluate_decisions thinking tool; Jev complements the generative model rather than replacing it. Per the yibie list, from version 1.12 the Kernel picks a specialist subagent by Choice, and retrieved RAG passages are withheld from the generating model when a prompt-injection Noul exceeds 0.70.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which team should handle this ticket? | — |
| Choice | Which specialist subagent should take this step? | — |
| Yes / no | Does this retrieved passage contain a prompt injection? | yes · no |
How it works
- Install with the typesafe extra and set TYPESAFE_API_KEY
- Agents call self.decisions.evaluate with state and questions
- Kernel routes steps to specialists by Choice
- RAG passages above 0.70 injection probability are withheld
- Generative model handles planning and execution
Strengths
- Decision client is separate from the text model
- Injection screening on retrieved passages before generation
- Durable, orchestrator-free multi-agent execution
Caveats
- Jev is one feature in a large framework
- Threshold of 0.70 is a fixed default per the list
Getting started
`pip install "jarviscore-framework[typesafe]"` and set TYPESAFE_API_KEY.At a glance
- Type
- Library
- Status
- production
- Language
- Python
- Runtime
- multi-agent framework
- Has code
- yes
Use cases and integrations
How curated lists describe it
Agent frameworks: Python multi-agent runtime that ships Jev natively from 1.12, where agents ask typed `Choice`, `Score` and `Noul` questions through a decision client separate from the text model, the Kernel picks a specialist subagent by `Choice`, and each retrieved RAG passage is withheld from the generating model when its prompt-injection `Noul` exceeds 0.70.
yibie/awesome-jev
Frequently asked questions
How does JarvisCore use Jev?
Agents ask Choice, Score and Noul questions through an injected decision client; the Kernel routes by Choice and RAG passages are screened by Noul.
Does Jev replace the LLM?
No. The README says Jev complements the generative model used for planning and execution.
How do I enable it?
Install the typesafe extra and set TYPESAFE_API_KEY.
Related projects
Composio agent SDK with an optional TypeSafe provider that uses Jev to shortlist and pick tools and closed-set arguments.
ProjectproductionTypeScriptOpen-source agent guardrails that use Jev or LAYA to check prompts, retrieved content, tool calls and outputs for injection and risk.
ProjectbetaPythonMiddleware for LangChain deepagents where Jev ranks and verifies which SKILL.md skills each turn needs, keeping the rest out of the prompt.
Agent skillbetaPythonDecision middleware for AI agents that uses local Laya to pick tools and judge results, falling back to an LLM when unsure.
ProjectbetaPythonSix-file DSPy coding agent where Jev answers closed questions (safety, tool fit, history, delegation) and Python owns policy.
ProjectprototypePythonAgent tool router that keeps obvious calls on a fast path, uses Jev via OpenRouter for ambiguous tool choices, and offers bounded MCTS.
MCP serverprototypePython