Skip to content

Back to the catalog

merijjeyn/jive
github.comrepo

merijjeyn/jive

Jive is an open-source agent framework that replaces the standard LLM-call/tool-call loop with 'Graph Calls' — DAG-based workflows combining traditional tool calls with fast 'Jev' (System One) model calls for repetitive or intuitive steps. It aims to reduce token cost and latency versus agents like Claude Code and Codex by only invoking expensive reasoning LLM calls when needed, letting cheaper System One models follow through on planned execution graphs.

TL;DR Open-source agent framework using DAG 'Graph Calls' to mix slow LLM reasoning with fast System One model execution for cheaper, faster agentic loops.
Open original ↗
Document

What it claims

Jive is a working, installable agent CLI (not a paper) built around "Graph Calls" — the summary covers the concept, but the page adds the actual install path, design constraints, and the lineage it builds on. It positions itself as "Agent 2.0," carrying forward the pi coding agent scaffold (see pi.dev): minimal, customizable, no MCP, no sub-agents — the author's stated view is that "clis are enough."

How to use it

One-line install:

curl -fsSL https://raw.githubusercontent.com/merijjeyn/jive/main/install.sh | sh

The docs it advertises cover: quick start and project configuration, sessions, headless commands, skills, extractors, a graph contract (the graph language the planner writes), planner context/compaction and Jev input limits, and a taskground for evaluating the planner. Contributors are asked to add their test tasks back to the repo to build a shared open-source eval dataset.

Design points the summary omits

  • No MCP, no multi-agent architecture — deliberately. The author argues fast/slow thinking should be a native primitive (System One models like Jev) rather than emulated by workflows or agent swarms.
  • Graphs are framed as a "higher-level programming layer" for the agent, explicitly not fixed workflow management — they stay flexible per-task.
  • Claims cache-efficient, token-efficient, eager execution with low resource use; works with both System One and System Two models.
  • Named target use cases beyond coding: customer support, lawyer-facing assistants, internal analytics agents — enterprise latency/cost reduction "without compromising quality."
  • Always open source and free.

Limits

The efficiency-vs-Codex/Claude-Code comparison is asserted by the author ("as you can see") with no benchmark numbers, baselines, or task settings reproduced on the page — treat the speed/token claims as unverified. The author concedes Jive is worse for "certain tasks" where an agent aligned with its training data wins, and stakes the whole bet on a claimed correlation between model intelligence and Jive's effectiveness — i.e. the payoff is forward-looking ("as models get better... the gap will be undeniable"), not established today. The author himself hedges: "I'm not sure if this is it." No licence file, version, or maturity signal is stated beyond "open source and free."

Organisations named here

  • Claude Code
  • Jev

Technologies named here

  • Jev
  • Jive
  • Codex
  • Claude Code
  • Pi coding agent
  • Taskground
  • Pi.dev

Techniques named here

  • Graph Calls
  • Native fast/slow agentic thinking

Related in agents