indusagi

A terminal-first AI coding-agent framework for Python. A zero-SDK LLM gateway, a pure-FSM agent runtime, a built-in tool suite, MCP interop, a multi-agent crew layer, and a Textual terminal UI — all in one importable package with a py.typed marker and no heavy dependencies in the core.

pip install indusagi

For the interactive terminal UI and MCP server mounting, install the extras:

pip install "indusagi[mcp,tui]"

The core (gateway, runtime, tools, print/wire CLI modes) needs only httpx, pydantic, wcwidth, python-ulid, and regex. The tui extra adds Textual, Pygments, and markdown-it-py; the mcp extra adds the official mcp SDK.

Requires Python 3.11+.

Use it as a library

import asyncio
from indusagi.agent import Agent

async def main():
    agent = Agent(model="claude-sonnet-4-5")   # reads ANTHROPIC_API_KEY from env
    reply = await agent.prompt("say hello in five words")
    print(reply)

asyncio.run(main())

Or talk to the model gateway directly (zero provider SDKs, just httpx wire connectors):

from indusagi.llmgateway import complete, get_card

reply = await complete("claude-sonnet-4-5", conversation, options)
card = get_card("claude-sonnet-4-5")          # context window, pricing, etc.

Every subsystem is importable on its own:

import indusagi.runtime        # the agent FSM, sessions, compaction
import indusagi.capabilities   # the built-in tool suite
import indusagi.interop        # MCP client/server bridge
import indusagi.swarm          # file-backed multi-agent crews
import indusagi.smithy         # the agent-builder

Use it as a CLI

The package installs the pindusagi console script:

pindusagi --help                      # usage
pindusagi --version                   # indusagi 0.1.2
pindusagi -p "explain this repo"      # one-shot: print the answer and exit
pindusagi --json                      # NDJSON wire protocol over stdio (for hosts)
pindusagi                             # interactive Textual TUI (needs the [tui] extra)
pindusagi auth login                  # OAuth + PKCE login
pindusagi auth status                 # show stored credentials

Flags include --model/-m, --interactive/-i, --cwd, --system, --no-tools, and repeatable --mcp <server>. With a prompt and no TTY it answers once and exits; attended with no prompt it opens the interactive loop. API keys come from the usual provider env vars (ANTHROPIC_API_KEY, OPENAI_API_KEY, …); state lives under ~/.pindusagi/ (override with INDUSAGI_HOME).

The TUI provides streaming markdown, structured diffs and tool cards, twelve dialogs (model / session / theme / login / …), a footer with token/cost/context stats, Esc to abort a run, and Ctrl-C to exit.

What's inside

indusagi/
├── llmgateway/     # zero-SDK LLM provider gateway (httpx wire connectors)
├── runtime/        # pure-FSM agent runtime: reducer, conductor, sessions
├── capabilities/   # built-in tool suite over Fs/Shell seams
├── interop/        # MCP bridge (official `mcp` SDK)
├── connectors/     # Composio SaaS connectors
├── swarm/          # file-backed multi-agent crew layer
├── smithy/         # agent-builder meta-tool
├── tracing/        # observability: segments, sinks, redaction, sampling
├── shell_app/      # CLI: print / NDJSON wire / REPL runners, OAuth login
├── tui/            # terminal primitives: keys, width, fuzzy, keybindings
├── react_ink/      # Textual widgets (markdown stream, diff view, dialogs)
├── ui_bridge/      # runtime ↔ TUI projection + interactive Textual app
└── ai/ agent/ mcp/ memory   # high-level facades over the core

849 models across 24 providers ship in the catalog (indusagi.llmgateway): Anthropic, OpenAI, Google, Bedrock, NVIDIA, Kimi, Ollama, OpenRouter, and more.

Supported providers

API keys are read from the standard env vars. Anthropic, OpenAI, Google (Generative AI + Vertex), Azure OpenAI, NVIDIA, Kimi, and Ollama have first-party wire connectors; hundreds more models route through the OpenAI-compatible connector. OAuth login is supported for Anthropic and OpenAI.

Development

python3.11 -m venv .venv
.venv/bin/pip install -e ".[dev,mcp,tui]"
.venv/bin/pytest            # 1,349 tests; network-free (mock connector, respx, Textual Pilot)

The package ships py.typed, the model-catalog data, and the agent-builder knowledge pack inside the wheel.

License

MIT. See NOTICE and CREDITS.md.