jsat 0.1.0

JaySoft AI Tools — open-source codebase intelligence shell and SDK

JSAT — JaySoft AI Tools

Codebase intelligence shell and SDK. Lightweight by default.

What is JSAT?

JSAT gives any AI (Claude, Codex, Gemini, or local Ollama) deep, structured understanding of your codebase — so you spend less time explaining context and more time shipping.

• Interactive shell like IPython, but for codebase intelligence
• Python SDK — from jsat import JSAT
• Works offline with local Ollama — no API keys required
• Lightweight by default — pip install jsat is ~80MB and starts in <800ms

Quick Start

pip install jsat          # core only
pip install jsat[local]   # + Ollama for local AI (recommended)

cd your-project/
jsat init --profile solo  # write .jsat.yaml
jsat index .              # build the codebase graph
jsat                      # start the shell

JSAT Shell v0.1.0
> what does this project do?
> which services write to the orders table?
> blast-radius src/payment/refund.py

Installation Profiles

Profile	Command	Size	Use case
Core	pip install jsat	~80MB	Quick evaluation
Local AI	pip install jsat[local]	~85MB	Solo dev with Ollama
Standard	pip install jsat[standard]	~120MB	Individual engineers
Team	pip install jsat[team]	~200MB	Engineering teams
CI	pip install jsat[ci]	~90MB	CI pipelines (no AI cost)
Full	pip install jsat[all]	~350MB	Power users

System Auto-Detection

JSAT probes your machine on first run and selects the right backends:

jsat doctor   # see what's detected and what to install

Your setup	Graph	Embeddings	Cache
Laptop, Ollama running	SQLite	nomic-embed-code (local)	disk
Team server (Neo4j+Qdrant+Redis)	Neo4j	text-embedding-3-small	Redis
ARM (Apple M2, Pi)	SQLite	nomic-embed-code via Metal	disk
CI environment	SQLite	none (skipped)	memory

Python SDK

from jsat import JSAT

js = JSAT(repo=".")
js.index()

# Natural language query
result = js.query("what calls the refund endpoint?")
print(result.answer)

# Blast radius — trace impact of a change
report = js.blast_radius("src/payment/refund.py")
for impact in report.impacts:
    print(f"{impact.severity}: {impact.node_name}")

# Incident investigation
incident = js.investigate_incident("500 errors on checkout since 14:00")
for h in incident.hypotheses:
    print(f"Score {h.score}: {h.commit_summary}")

# Export for sharing
js.export("backup.jsat.zip")

CLI Reference

jsat index [PATH] [--branch HEAD] [--force]
jsat shell
jsat doctor [--refresh] [--json]
jsat init --profile solo|team|ci|raspberry-pi
jsat export OUTPUT
jsat import ARCHIVE
jsat skills list
jsat version

Architecture

JSAT is built around a graph-native architecture:

• tree-sitter for multi-language AST parsing
• SQLite (local) or Neo4j (team) for the codebase graph
• Ollama (local) or any cloud LLM for AI features
• Lazy loading — nothing heavy loads until you use it

All backends implement ABCs — tools never call backends directly.

Project

• Author: Jay Prakash Sonkar (@iamjpsonkar)
• Email: [e-mail]
• License: MIT
• Specification: plan.md | Implementation guide: prompt.md