soothe-nano 1.0.6

Soothe Nano — batteries-included SootheNanoAgent on deepagents

soothe-nano

A ready-to-run coding agent you can drop into a script, CLI, or app.

Built on soothe-deepagents (filesystem, shell, subagents, skills, MCP). Nano adds the pieces you usually wire yourself: workspace safety, progressive tools/skills, research subagents, and a config-driven factory.

Vision

Give builders a production-shaped SootheNanoAgent in a few lines of Python.

• Start small — chat, tools, or full composition
• Stay portable — embed in notebooks, CLIs, or your own service
• Compose what you need — tools, memory, subagents, skills, and MCP via config

Architecture

soothe-deepagents   agent harness (FS, shell, memory, skills base)
        ↓
soothe-sdk          shared contracts (protocols, events)
        ↓
soothe-nano         SootheNanoAgent + toolkits + subagents + MCP

create_nano_agent(config)
        │
        ├─ model + middleware stack
        ├─ tools (builtin groups + yours)
        ├─ subagents (planner, research, browser, …)
        ├─ skills (progressive discovery)
        └─ MCP (on-demand activation)

Features

Area	What nano provides
Tools	Builtin groups: shell, file ops, HTTP, search, data, …
Subagents	Ready: plan, deep/academic research, browser
Skills / tools in context	Progressive loading — activate what the turn needs
Workspace	Scoped workspace + security defaults
Config	YAML / SootheConfig factory
Memory	Optional long-term memory via protocols
MCP	Registry and on-demand adapters

vs deepagents

	deepagents	soothe-nano
What you get	Opinionated harness	Harness plus coding product defaults
Tools	Bring your own	Builtin groups out of the box
Subagents	You define them	Ready plan / research / browser
Skills / tools in context	Base support	Progressive loading
Workspace	Pluggable backends	Scoped workspace + security defaults
Config	Code-first	YAML / SootheConfig factory

Use deepagents when you want a minimal harness and full control.
Use nano when you want a coding agent that already knows how to work in a repo.

When to use nano

Scenario	Fit
Coding assistant in a repo	✅ Files, shell, plan out of the box
Research / browsing agent	✅ Deep research, academic, browser subagents
Embed in your product	✅ Library API, no daemon required
One-shot / headless CLI	✅ See fj-ai
Plugin / toolkit author	✅ Depends on nano only
Simple Q&A chat	✅ Strip tools/subagents as needed

Install

uv add soothe-nano

Quick start

from soothe_nano import create_nano_agent
from soothe_nano.config import SootheConfig

agent = create_nano_agent(SootheConfig())
# agent.ainvoke / streaming — see examples/

Library examples live in examples/:

1. Pure model (no tools)
2. With tools
3. With memory
4. With subagents
5. Full composition

python packages/soothe-nano/examples/01_pure_nano_example.py

Productive CLI (reference: fj-ai)

fj-ai is a production one-shot coding CLI built only on soothe-nano (no soothe daemon). Use it as the integration blueprint:

pip install fj-ai   # or: uv tool install fj-ai
fj explain this repo
fj -f what did we decide last time?

Integration pattern

A headless CLI typically does four things on top of nano:

1. Load config — ~/.soothe/config/nano.yml, or zero-config from OPENAI_API_KEY / ANTHROPIC_API_KEY
2. Force SQLite for standalone runs (threads survive across process exits)
3. Build the agent with create_nano_agent, pin workspace, attach a checkpointer
4. Stream agent.astream(...) and close the aiosqlite connection on exit

Minimal sketch (same shape as fj_ai/agent.py):

from __future__ import annotations

import asyncio
from contextlib import asynccontextmanager
from pathlib import Path
from typing import Any, AsyncIterator

from soothe_nano import SootheNanoAgent, create_nano_agent
from soothe_nano.config import SOOTHE_HOME, SootheConfig
from soothe_nano.resolve import resolve_checkpointer


def load_config(path: Path | None = None) -> SootheConfig:
    cfg = path or (SOOTHE_HOME / "config" / "nano.yml")
    if cfg.is_file():
        return SootheConfig.from_yaml_file(str(cfg))
    return SootheConfig()  # OPENAI_API_KEY / ANTHROPIC_API_KEY


def apply_cli_defaults(config: SootheConfig) -> SootheConfig:
    durability = config.agent.protocols.durability.model_copy(
        update={"backend": "sqlite", "checkpointer": "sqlite"}
    )
    protocols = config.agent.protocols.model_copy(update={"durability": durability})
    agent = config.agent.model_copy(update={"protocols": protocols})
    persistence = config.persistence.model_copy(update={"default_backend": "sqlite"})
    return config.model_copy(update={"agent": agent, "persistence": persistence})


@asynccontextmanager
async def open_sqlite_checkpointer(config: SootheConfig) -> AsyncIterator[Any | None]:
    result = resolve_checkpointer(apply_cli_defaults(config))
    db_path = result[1] if isinstance(result, tuple) and isinstance(result[1], str) else None
    if not db_path:
        yield None
        return

    import aiosqlite
    from langgraph.checkpoint.sqlite.aio import AsyncSqliteSaver
    from soothe_sdk.utils.serde import create_soothe_serde

    Path(db_path).parent.mkdir(parents=True, exist_ok=True)
    conn = await aiosqlite.connect(db_path)
    checkpointer = AsyncSqliteSaver(conn, serde=create_soothe_serde())
    await checkpointer.setup()
    try:
        yield checkpointer
    finally:
        await conn.close()  # required — aiosqlite uses a non-daemon thread


async def build_agent(
    config: SootheConfig,
    *,
    checkpointer: Any | None = None,
) -> SootheNanoAgent:
    agent = create_nano_agent(apply_cli_defaults(config))
    if checkpointer is not None:
        agent.graph.checkpointer = checkpointer
    return agent


async def run_once(query: str, *, thread_id: str) -> None:
    config = load_config()
    async with open_sqlite_checkpointer(config) as checkpointer:
        agent = await build_agent(config, checkpointer=checkpointer)
        async for chunk in agent.astream(
            query,
            config={"configurable": {"thread_id": thread_id}},
            stream_mode=["messages", "updates", "custom"],
        ):
            # render tokens / tool progress to stdout (see fj_ai/stream.py)
            _ = chunk


if __name__ == "__main__":
    asyncio.run(run_once("summarize this repo", thread_id="cli-demo"))

What fj adds on top of nano

Concern	fj-ai approach
UX	One-shot fj <query…>; -f / -t resume threads; -l list
Persistence	SQLite checkpointer under $SOOTHE_DATA_DIR
Skills	Package builtin_skills/ via register_builtin_skill_root
Workspace	SOOTHE_WORKSPACE / cwd for file + shell tools
Streaming	Quiet progress line + full final answer (stream.py)
Setup	fj setup writes ~/.soothe/config/nano.yml

For a full TUI / StrangeLoop host on the same stack, see mirasoth/soothe. For the slim CLI product, clone caesar0301/fj-ai.

Package layout

soothe_nano/
  agent/       SootheNanoAgent, create_nano_agent
  config/      SootheConfig
  toolkits/    Builtin tool groups
  subagents/   plan, research, browser, …
  middleware/  Progressive tools/skills, workspace, policy
  skills/      Catalog + progressive search
  mcp/         MCP registry / adapters
  backends/    Persistence helpers

Development

From packages/soothe-nano/:

make help              # list targets
make sync-dev          # sync deps
make format lint       # format + lint
make test-unit         # unit tests
make test-integration  # integration tests (--run-integration)
make examples          # run examples
make build             # build dist/