Nehanda CLI

A local AI coding substrate powered by the Nehanda fine-tuned model. Free and open-source.

Overview

Nehanda CLI is a local AI coding agent that brings the Nehanda model into your terminal. It is a fork of aimee — wiring Nehanda in as the primary reasoning model: a fine-tuned Qwen3.6 27B multimodal stack for technical writing, deep research, coding, document review, and vision.

No Docker. Three native binaries + postgres + local embedder.

Architecture

  Your Terminal
    └─ aichat (TUI) ──────────────────────────→ nehanda-server :8740 → nehanda-server (UDS)
  • 4-tier memory compaction (~86% token reduction pre-egress)
  • Code index + KB — postgres + pgvector, never leaves machine
  • Supervisor/Delegate task fan-out → local/LAN Ollama (free)
          │  compacted payload only
          ▼
  Nehanda 27B on EC2 (nehanda.asoba.co:8000)
  • vLLM, tensor-parallel
  • nehanda-rag-synthesis-27b

The stack runs entirely on your machine except for the Nehanda model inference, which hits a remote EC2 endpoint. Only the compacted payload leaves your machine — your code and knowledge base stay local.

Quick Start

Requirements

macOS or Linux
cmake 3.16+, make, git, C11 compiler
Homebrew (macOS) or apt/dnf (Linux)
Ollama (optional — for local and LAN delegate workers)

Install

git clone https://github.com/AsobaCloud/nehanda-cli.git
cd nehanda-cli
./install.sh
nehanda

install.sh builds the nehanda stack and installs aichat as the interactive TUI. It handles postgres, pgvector, and all build dependencies — one command, no follow-up steps.

./install.sh    # ~40s on re-run with cached build
nehanda         # interactive TUI
nehanda chat "…"  # one-shot message (server stream, no TUI)

Add to your shell profile:

export PATH="$HOME/.local/bin:$PATH"
export PATH="/opt/homebrew/opt/postgresql@17/bin:$PATH"   # macOS only

Automatic Workspace Registration

When you run nehanda from any git repository, it automatically adds that directory as a workspace. This enables the agent to access files from your actual working directory instead of being sandboxed in worktrees.

Key Features

4-Tier Memory Compaction

Every session starts already knowing what the last one learned. A curator pipeline distills raw sessions into a typed knowledge base, dedupes facts, flags contradictions, and lets stale detail decay. Approximately 86% token reduction before any payload reaches the model.

Codebase as a Map

Nehanda CLI indexes your code into a symbol and call graph. The agent finds callers, traces an edit’s blast radius before it writes a line, and works from the graph instead of reading your files over again each session. The graph spans repositories — one dependency map covers every repo you work across.

Supervisor/Delegate Task Fan-Out

Send summarization, review, and boilerplate to the cheapest model that can handle it. Local models on your GPU via Ollama cost nothing. The primary agent gets a compact result back instead of raw content, saving on both the delegate and the primary’s context.

Guardrails and Isolation

Sensitive files like .env, private keys, and production configs are blocked before the agent can touch them. Each session runs in its own git worktree, so two sessions never step on each other.

Slash Commands

Command	Description
`/model`	Change agent model or orchestrator model
`/bearer`	Update bearer token in config atomically
`/token`	Last 24h token usage breakdown
`/stats`	KB, memory, token usage, and service health
`/auth`	Auth status / login stub
`/config`	Show current config
`/clear`	Start a new conversation
`/help`	List all commands
`/exit`	Quit

Self-Hosting

Nehanda CLI supports running with a local model instead of the EC2 endpoint. Any OpenAI-compatible model can be configured as the primary agent or orchestrator, including models from ZAI, OpenAI, or self-hosted models. See the self-hosting guide in the repo.

Model Configuration

Nehanda-cli supports any OpenAI-compatible model as your primary agent or orchestrator. You can configure models through the agent configuration system, allowing you to use models from various providers without being locked to a specific provider.