Pawan

पवन pawan

Self-healing CLI coding agent

Rust-native. 34 tools. 18 commands. No subscription.

Get Started GitHub

Self-Healing

pawan heal fixes compilation errors, clippy warnings, and failing tests automatically.

AI Commit

pawan commit -a stages files, generates a message from the diff, confirms and commits.

Code Review

pawan review reviews your changes with severity levels and actionable fixes.

Watch Mode

pawan watch polls your project and auto-heals on errors. Set it and forget it.

Headless

pawan run "prompt" executes tasks non-interactively for scripting and CI.

Streaming TUI

Markdown rendering, vim keybindings, live token streaming, tool progress indicators.

Compiler as Auditor

Every iteration is type-checked. cargo check runs in the loop — the Rust compiler catches hallucinations before they land.

Local Inference

MLX on Apple Silicon, Ollama on Linux, llama.cpp via lancor anywhere. $0/token. Falls back to NIM cloud automatically.

Why Pawan

Born from a late-night revelation: CLI coding agents are superpowers. Run 8 in parallel, each on different branches, merge their work. We reverse-engineered the best, then built our own in Rust — integrated with our AI runtime, self-hosted, zero subscription.

"U CAN RUN 8 OF THEM IN PARALLEL WITH EASE IN THE SAME TERMINAL WINDOW"
— the moment it clicked

Named after Power Star Pawan Kalyan — black belt martial artist, Telugu cinema cult icon, Deputy CM of Andhra Pradesh. A man who went from action movies to running a state government. That energy: raw power, cult following, fearless execution. पवन means "wind" in Sanskrit, and the name was given to Kalyan by a karate association after a public martial arts demonstration. From powercode to pawan code to just pawan — born in a dirmacs WhatsApp jam session, built by models operating inside terminal agent harnesses, powered by NVIDIA NIM.

dirmacs ecosystem

Pawan is one layer of a self-hosted AI infrastructure stack. Every component is open-source, MIT-licensed, and independently usable.

ares-server

LLM runtime — multi-provider routing, tool coordination, RAG, and agent orchestration as a self-hosted API server.

eruka

Context memory engine — a structured knowledge graph your agents can read, write, and query to stay oriented across sessions.

daedra

Web search MCP server — 7 backends, automatic fallback, stdio transport. Drop-in search for any MCP-compatible agent.

deagle

Code intelligence — tree-sitter + SQLite graph, BM25 keyword search, AST pattern matching. 9 languages, embeddable as a library.

thulp

Skill framework — typed tool definitions, SKILL.md parsing, multi-step skill execution, distillation and refinement via thulpoff.

aegis

Config manager — declarative system configuration for NIM models and agent settings, with drift detection and auto-sync.

Install

cargo install pawan
export NVIDIA_API_KEY=nvapi-...
pawan doctor  # verify setup

# Or go local ($0/token):
# MLX on Mac:  PAWAN_PROVIDER=mlx pawan
# Ollama:      PAWAN_PROVIDER=ollama PAWAN_MODEL=llama3.2 pawan
# llama.cpp:   cargo install pawan --features lancor
#              PAWAN_PROVIDER=lancor PAWAN_MODEL=./model.gguf pawan