tantra

The agent turn loop as a library. Install name tantra-harness, import name tantra.

What it is

• The harness owns the loop — sampling, tool dispatch, permissions, approval/suspend-resume, subagents, compaction, persistence. You supply an Agent (a declarative class of values and function references) and run turns against it.
• Tools, hooks and skills plug in. An Agent declares tools; a Harness takes hooks and a Skills source.
• Agents hold no live I/O. A resume can run in another process and look the agent up by name from the persisted session header.
• Configuration is explicit construction: tools are built by factories (web_search(api_key=...)) in your own module, so keys are per-agent and a misconfiguration fails at construction, not mid-turn. The library never reads the environment.
• Shipped tools live under tantra.extratools.* and their dependencies are extras, so the base install stays light — pydantic, httpx, openai only. The shell tools are stdlib and need no extra.

Install

Import-name collision. The unrelated PyPI project tantra also installs an import tantra. Installing both into one environment clobbers the import silently. Do not co-install them.

Basic usage

import asyncio

from tantra import Agent, Harness, OpenAICompatible, SQLiteStore
from tantra.extratools.doc import read_doc
from tantra.extratools.shell import ShellGuard, bash
from tantra.extratools.web import web_fetch, web_search
from tantra.providers.base import TextDelta


class Researcher(Agent):
    model = "gpt-5"
    prompt = "You answer questions using the web, local documents and the shell."
    tools = [bash(), web_search(api_key=BRAVE_API_KEY), web_fetch(), read_doc()]
    permissions = {"bash": "allow"}


async def main() -> None:
    store = SQLiteStore("sessions.db")
    await store.setup()
    harness = Harness(
        OpenAICompatible("https://api.openai.com/v1", OPENAI_API_KEY),
        store,
        [Researcher],
        hooks=[ShellGuard()],
    )
    session = await harness.create_session(Researcher)
    async for emitted in harness.run(session.id, "What changed in Python 3.13?"):
        if isinstance(emitted.event, TextDelta):
            print(emitted.event.text, end="", flush=True)


asyncio.run(main())

Notes on the snippet:

• harness.run(...) is an async iterator of Emitted. The turn only advances while the stream is consumed; tantra.collect drains it into a list.
• bash declares permission="ask", so a headless run must override it — hence permissions = {"bash": "allow"}. Globs work ("web_*": "allow").
• ShellGuard() denies destructive commands with a reason the model sees. ShellGuard(on_trip="ask") escalates to the human approval flow instead. It is a guardrail, not a sandbox.

Docs

• https://malayh.github.io/tantra/
• https://malayh.github.io/tantra/docs/

Reference apps

• apps/sarthi in apps/sarathi — a fully usable Perplexity clone with multi tanancy, memoery, parallel agents, web search built in
• apps/agni in apps/agni — a terminal coding agent built on this library: REPL, permission prompts, skills, memory, compaction.

License

Apache-2.0