Skip to content

Agents that survive production

JarvisCore is an open source Python framework for building autonomous multi-agent AI systems. Most agent frameworks get you a demo; JarvisCore gets you an operator: agents that run unattended for weeks, remember last month, fail loudly, and leave a flight record you can read. We run our own agents on it, with real budgets. Every hardening release comes from those scars.


Why JarvisCore

Built by running our own agents unattended, with real budgets. Four rules fell out of that, and they are enforced in code, not promised in prose:

  • Honest context. Nothing is truncated silently. Clips are labeled, originals archived, summaries say what they summarize.
  • Loud failures. Failed steps say so. Partial work survives. Rate-limit storms are absorbed, not crashed on.
  • Two profiles, no mushy middle. CustomAgent: you bring the brain. AutoAgent: the full cognitive stack. It writes its own integrations, repairs them in a sandbox, and banks verified work for reuse.
  • No single point of death. Agents coordinate over a SWIM gossip mesh. Memory compounds across sessions via Athena. Credentials stay out of agent reasoning via Nexus.

Here is what those rules buy you in practice.


What JarvisCore provides

Agent Profiles

Two execution models

AutoAgent runs a full cognitive loop internally: observe, orient, decide, act. CustomAgent exposes the execution loop directly for deterministic control. Both share the same infrastructure.

Memory

Four-tier agent memory

Working scratchpad, episodic ledger, LLM-compressed long-term summaries, and optional cross-session semantic memory via Athena MemOS. Wired into AutoAgent automatically when ATHENA_URL is set. Context that survives restarts and compounds across weeks.

Communication

Self-organising agent mesh

Agents discover and message each other via a PeerClient API over SWIM gossip and ZMQ. No central orchestrator to die: nodes join, fail, and rejoin. Identical code on a single process or across distributed machines.

Integrations

46 service integrations

Slack, GitHub, Zoom, SAP, NetSuite, MS Graph, Salesforce, and 40 more. 237+ prebuilt actions your agents can call directly. No glue code, no auth wiring.

Browse integrations →

Auth

Nexus credential layer

Agents call third-party APIs without ever touching raw credentials. OAuth2, API keys, and basic auth are all managed by Nexus and kept out of agent reasoning.

Nexus guide →

Observability

Full-stack tracing

Every agent turn, tool call, and LLM request is traced automatically. Redis PubSub for live streams, JSONL for compliance, Prometheus for operational dashboards.

Observability guide →

Control

Human-in-the-loop

HITLQueue intercepts decisions that exceed confidence thresholds, routes them to a review inbox, and resumes execution once a human responds. First-class, not an afterthought.

HITL guide →


Quickstart

One LLM provider key is the only required configuration. Install, initialise, and run your first autonomous agent:

Install & initialise
pip install jarviscore-framework
jarviscore init
cp .env.example .env   # add your LLM key
jarviscore check        # verify dependencies
main.py
import asyncio
from jarviscore import Mesh, AutoAgent


class ResearcherAgent(AutoAgent):
    name = "Researcher"
    role = "researcher"
    system_prompt = "You are a rigorous research analyst."


async def main():
    mesh = Mesh()
    mesh.add(ResearcherAgent)
    await mesh.start()
    result = await mesh.run_task(
        agent="researcher",
        task="What are the key architectural trade-offs in multi-agent systems?",
    )
    print(result)


asyncio.run(main())

Where to start

If you are new to JarvisCore, read in this order:

  1. Getting Started: install, configure, and run your first agent
  2. Architecture Overview: the mental model for how the framework fits together
  3. Agents: what an agent is, its identity and lifecycle
  4. Language Models: how JarvisCore uses multiple LLMs simultaneously
  5. Memory: how agents maintain and recover context
  6. Agent Personas: how profiles shape autonomous behaviour

If you are evaluating for a specific use case:


Explore the ecosystem

Reference Full API surface, configuration keys, and CLI flags: view reference
Source Browse the code, open issues, and submit PRs: GitHub
Community Questions, showcases, and early feature previews: Discord
Blog Engineering deep-dives and architecture walkthroughs: read the blog

Star us on GitHub

Help more developers discover JarvisCore. Every star makes us easier to find and keeps the project growing.

Join the community

Ask questions, share what you're building, and get early previews of new features. Come say hi.