Jiva

Jiva is an autonomous AI agent for the terminal and cloud. It works with any OpenAI-compatible model provider and supports two execution modes — a general-purpose two-agent system (Manager + Worker) for complex tasks, and a code-optimized single-loop engine with LSP integration for software engineering.

Demo

---

Installation

npm install -g jiva-core
jiva setup

# Verify your model + config can do code mode well:
jiva benchmark

The setup wizard opens with a provider selection menu. Choose your provider and Jiva auto-fills the endpoint, model name, and tool format — you only supply your API key. Supported providers: Krutrim, Groq, Sarvam, OpenAI, and any OpenAI-compatible endpoint.

After setup, run jiva benchmark to confirm your chosen model and configuration work optimally in code mode. It runs the baseline (taskstore) suite — a deterministic set of coding tasks — against your config; a healthy setup passes all of them. See Benchmarking for tuning and capability comparison.

Development install

git clone https://github.com/KarmaloopAI/Jiva.git
cd Jiva
npm install && npm run build && npm link
jiva setup

---

Quick Start

# General-purpose interactive session
jiva chat

# Code mode — optimized for software engineering tasks
jiva chat --code

# Code mode with plan-then-approve flow
jiva chat --code --plan

# Single prompt
jiva run "What changed in the last 5 commits?"

# Single prompt in code mode
jiva run "Add error handling to the database module" --code

REPL commands

Command	Description
/help	Show available commands
/tools	List active tools
/servers	Show MCP server status
/save	Save current conversation
/load	Resume a saved conversation
/list	Browse all saved conversations
/reset	Clear conversation history
/<skill-name>	Load a skill (e.g. /graphify)
/exit	Exit Jiva

---

Execution Modes

General mode (default)

A two-agent pipeline designed for complex, multi-step tasks:

User Request
    ↓
Manager Agent    — plans, decomposes, and synthesizes
    ↓
Worker Agent     — executes subtasks with MCP tools
    ↓
Response

The Worker has access to any configured MCP server (filesystem, browser automation, shell commands, etc.). The Manager synthesizes results from all subtasks into a final coherent response. Simple conversational messages are answered directly without executing subtasks.

Code mode (--code)

A single streaming loop optimized for coding tasks:

User Request
    ↓
CodeAgent
    loop until done:
        LLM call with tool definitions
        ↓
        Tool execution (in-process, no subprocess)
        ↓
        LSP feedback after file edits
    ↓
Response

Code mode reduces latency by eliminating inter-agent overhead and running all file tools directly in the Node.js process. It includes a multi-strategy edit engine that reliably handles indentation drift and whitespace inconsistencies.

Available tools in code mode: