Dhara

A minimal, secure, language-agnostic coding agent harness.

Dhara (धारा, "flow" in Sanskrit) is the engine that powers AI coding agents.
It provides the agent loop, tool management, session persistence, and extension
protocol — without being tied to any LLM provider, UI framework, or programming
language.

npx @zosmaai/dhara "List the files in this project"

Features

• 🔄 Agent Loop — LLM → tool → LLM cycle with streaming, cancellation, and error recovery
• 🔌 Extension Protocol — JSON-RPC 2.0 wire protocol. Write extensions in any language
• 📦 20+ LLM Providers — OpenAI, Anthropic, Google, Mistral, Groq, DeepSeek, Bedrock, and more via pi-ai
• 🛡️ Capability Sandbox — Declare what tools need. Block everything else.
• 💾 Session Persistence — Append-only JSONL format. Resume conversations across restarts.
• 🖥️ TUI + REPL — Full-screen terminal UI with syntax highlighting, or line-based REPL
• 📝 Context Files — AGENTS.md / CLAUDE.md auto-loading per project
• 🔧 Extensions — Any language, any tool, any hook. Extensions are subprocesses, not function calls.

Quick Start

# Run with a prompt (auto-detects opencode-go, or use --provider)
npx @zosmaai/dhara "List the files in this project"

# Interactive mode (TUI)
npx @zosmaai/dhara

# Choose a provider
export GOOGLE_API_KEY="..."
npx @zosmaai/dhara --provider google --model gemini-2.5-flash

Providers

Dhara supports 20+ LLM providers through the pi-ai integration.
Set the corresponding environment variable and use --provider:

Built-in providers (no pi-ai needed): openai, anthropic, opencode-go.

Standard Tools

Dhara ships with 6 essential tools. Everything else is an extension.

Network tools (web fetch, web search), database tools, git operations — all belong in extensions.

Extensions

Extensions communicate via JSON-RPC 2.0 over stdin/stdout. Any language works.

# my-extension/main.py
import sys, json

for line in sys.stdin:
    req = json.loads(line)
    if req["method"] == "initialize":
        print(json.dumps({"jsonrpc":"2.0","result":{
            "protocolVersion":"0.1.0","name":"my-ext","version":"1.0.0",
            "tools":[{"name":"hello","description":"Say hello",
                      "parameters":{"type":"object","properties":{}}}]
        },"id":req["id"]}))
    elif req["method"] == "tools/execute":
        print(json.dumps({"jsonrpc":"2.0","result":{
            "content":[{"type":"text","text":"Hello from Python!"}]
        },"id":req["id"]}))
    sys.stdout.flush()

Install: copy to ~/.dhara/extensions/my-extension/ with a manifest.json.

Documentation

Full documentation is at docs.dhara.zosma.ai (coming soon).

• Getting Started Guide
• Writing Your First Extension
• Architecture Overview
• Extension Protocol Spec
• Session Format
• Roadmap

Architecture

┌─────────────────────────────────────────┐
│              ECOSYSTEM LAYER            │
│  Packages · Themes · Skills · Prompts   │
├─────────────────────────────────────────┤
│              EXTENSION LAYER            │
│  Tools · Providers · Renderers · Hooks  │
│  (any language, wire protocol)          │
├─────────────────────────────────────────┤
│               CORE LAYER               │
│  Agent Loop · Tool Interface · Sandbox  │
│  Session Format · Event Bus             │
│  (< 2,000 lines, no LLM, no UI)         │
└─────────────────────────────────────────┘

Project Status

Dhara is in active development. Phases 0–3 (Spec, Core, Standard Library, CLI) are
complete. Current work focuses on documentation, extension ecosystem, and launch prep.

See spec/roadmap.md for details.

Contributing

See CONTRIBUTING.md for guidelines.

License

MIT — see LICENSE. Spec documents are CC-BY-4.0 — see LICENSE-SPEC.

Built with ❤️ by Zosma AI.