Curry Leaves Agent Loop

A small, provider-agnostic, multi-agent kernel for building AI agents of any kind — in clean, readable Python.

Curry Leaves is a general-purpose agent kernel small enough to read in an afternoon. At its core is a streaming tool-use loop — call the model, run the tools it asks for, feed the results back, repeat — with everything a real agent needs built around it: sub-agents, skills, MCP tools, permission gating, session recording, automatic context compaction, and per-turn reasoning-effort sizing. Point it at any tools and any task; the kernel is domain-agnostic — it just happens to ship a batteries-included coding toolset and CLIs on top.

One engine, any provider, any UI. The loop knows nothing about a specific LLM wire format — each provider translates at its own boundary — so Anthropic, OpenAI, and Ollama are a config choice, not branching logic. Use it as a library (below), or launch either bundled CLI: a full-screen terminal UI or a line REPL.

from curry_leaves import Agent, Runner, coding_tools

agent  = Agent(model="claude-sonnet-4-5", tools=coding_tools())
result = await Runner(agent).run("Summarize README.md in three bullets.")
print(result.output_text)

---

Table of contents

• Features
• Install
• Quick start
• Configure a provider
• Terminal UI & REPL
• Library usage
• Environment variables
• Architecture
• Project layout
• Development
• Contributing
• Non-goals
• Acknowledgements
• License

Features

• Multi-provider. Anthropic, OpenAI, and Ollama (any OpenAI-compatible gateway) implement one Provider.stream() interface; the loop never changes. Streaming SSE is assembled into a neutral AssistantMessage at the provider boundary and nowhere else.
• Rich toolset. read, write, edit, find, search, bash, task_create/task_update/task_list/task_get, ask, current_time, web_fetch, web_search. Tool args are pydantic models → JSON Schema for free. Oversized output is offloaded to an artifact store the model can page through with read.
• MCP tools. Connect any Model Context Protocol server — stdio subprocess or HTTP — and pick specific tools by name with mcp_tools(); the result is a plain list[Tool] spliced into Agent(tools=[...]) like any preset. Servers can also be declared in settings.json's mcpServers key. MCP tools default to risk="exec", so they always go through the permission gate.
• Deferred tools + search_tools. Keep the advertised list lean; the model discovers more tools by keyword and activates them for the next turn — works across providers.
• Sub-agents. Declare subagents=[...]; the parent gets a task tool (delegation that returns a result) and a transfer tool (one-way handoff). Bounded recursion depth.
• Structured output. Give an agent an output_type (a pydantic model); the Runner injects the schema, validates the final reply, and retries on mismatch. result.output is typed.
• Auto-thinking. auto_thinking=True sizes reasoning effort (Anthropic thinking budget / OpenAI reasoning effort) per turn with a cheap classifier.
• Skills. Drop a SKILL.md under .curry-leaves/skills/<name>/; its teaser goes into the prompt and the model pulls the full body via read skill://<name> only when relevant (progressive disclosure).
• Permissions. An opt-in gate authorizes each tool call (allow / ask / deny), with standing approvals and contained-change auto-approval. Off by default — headless runs never hang.
• Sessions. Each run can be recorded to <home>/sessions/<id>/ (meta.json + transcript.jsonl). Branch a new session off an existing one with fork_session() (or /fork [n] in the CLIs) — it replays the recorded transcript into a fresh, independent session that diverges from there.
• Compaction. Long conversations are summarized as they near the context window — automatic, or on demand via Runner.compact() / the /compact command.
• Elision (opt-in). Before compaction ever fires, stale tool results (superseded reads, output untouched for N turns) are swapped for short stubs — originals preserved in the blob store, one read artifact://<id> away. Lossless, deterministic (no model call), and batched into sweeps so it never thrashes the provider's prompt cache. RunConfig(elision=ElisionConfig(enabled=True)).
• Typed, async, dependency-light. Python 3.11+, mypy --strict, ships py.typed. The library core rides on pydantic + httpx; the CLIs add Textual + Rich, and MCP support uses the official mcp SDK.

Install

pip install curry-leaves

Then launch a CLI:

curry-leaves        # full-screen terminal UI (alias: curry)
curry-leaves-repl   # line REPL (works with piped input)

Requires Python 3.11+.

Quick start

import asyncio
from curry_leaves import Agent, Runner, coding_tools

async def main() -> None:
    agent = Agent(
        model="claude-sonnet-4-5",
        instructions="You are a concise coding assistant.",
        tools=coding_tools(),
    )
    result = await Runner(agent).run("What does src/curry_leaves/runner.py do?")
    print(result.output_text)

asyncio.run(main())

Agent is a stateless definition (model, tools, instructions); Runner holds the live conversation and drives the streaming loop. Set an API key first (see below).

Configure a provider

export ANTHROPIC_API_KEY=sk-ant-...   # or
export OPENAI_API_KEY=sk-...

The provider is inferred from the model id (claude-* → Anthropic, gpt-*/o1-* → OpenAI, gemma*/llama*/qwen*/… → Ollama), or set CURRY_LEAVES_PROVIDER / pass an explicit provider to the Agent.

Local models via Ollama

Ollama speaks the OpenAI wire format, so any pulled tag works with no API key:

ollama pull qwen3
CURRY_LEAVES_MODEL=qwen3 curry-leaves-repl

from curry_leaves import Agent, Runner, coding_tools
agent = Agent(model="qwen3", tools=coding_tools())   # provider → Ollama

Set OLLAMA_HOST to point at a non-default server (http://host:port). Tool use needs a model with the tools capability (qwen3, llama3.x, gemma3, …). Reasoning-effort knobs are dropped for Ollama, and usage cost is $0 (local). You can point the same OpenAIProvider at any OpenAI-compatible gateway via OPENAI_BASE_URL or explicit provider options.

Terminal UI & REPL

A full-screen terminal UI (built on Textual) ships with the package — a header bar, a streaming transcript, live thinking blocks, spinner-tracked tool calls, indented sub-agent activity, a status bar, and a persistent input box. Finished turns land in real terminal scrollback; the active turn streams in place.

curry-leaves

╭──────────────────────────────────────────────────────────╮
│ curry-leaves · claude-sonnet-4-5 (anthropic)             │
│ /repo · 11 tools · subagents: explore, plan              │
╰──────────────────────────────────────────────────────────╯

you › what does src/curry_leaves/runner.py do?
ai ›
  → read({"path":"src/curry_leaves/runner.py"})
      1  """The Runner — holds one live conversation … """
The Runner composes an Agent with conversation state and drives the loop …

 ● ready                                    in 4213 · out 187 · $0.0155
╭──────────────────────────────────────────────────────────╮
│ › ask anything — /help for commands                      │
╰──────────────────────────────────────────────────────────╯

Slash commands: /help, /reset, /fork [n] (branch a new session from this conversation, optionally up through the nth user turn), /tools, /skills, /model, /stats, /clear, /compact [focus], /auto (toggle contained-change auto-approve), /autonomous (toggle self-drive mode), /exit. The TUI needs an interactive terminal (a TTY).

The line-streaming REPL is available as curry-leaves-repl — handy for piped / non-TTY input.

Library usage

Stream events instead of awaiting

async for event in runner.stream("Refactor the parser"):
    if event.type == "message_update" and event.delta and event.delta.kind == "text":
        print(event.delta.value, end="", flush=True)

Structured output

import pydantic
from curry_leaves import Agent, Runner

class Summary(pydantic.BaseModel):
    title: str
    bullets: list[str]

agent = Agent(model="claude-sonnet-4-5", output_type=Summary)

result = await Runner(agent).run("Summarize this repo.")
report = result.output   # validated Summary instance; the Runner retries on a schema mismatch

Sub-agents

from curry_leaves import Agent, Runner, coding_tools, explore_agent, plan_agent

agent = Agent(
    model="claude-sonnet-4-5",
    tools=coding_tools(),
    subagents=[explore_agent("claude-sonnet-4-5"), plan_agent("claude-sonnet-4-5")],
)
# The parent can `task` (delegate → result) or `transfer` (one-way handoff) to these.

MCP tools

from curry_leaves import Agent, Runner, coding_tools
from curry_leaves.mcp import McpServerStdio, mcp_tools

async with McpServerStdio(name="github", command="npx", args=["-y", "@modelcontextprotocol/server-github"]) as gh:
    agent = Agent(
        model="claude-sonnet-4-5",
        tools=[*coding_tools(), *await mcp_tools(gh, "search_issues")],
    )
    result = await Runner(agent).run("Find open issues about streaming.")

Or declare servers once in .curry-leaves/settings.json under mcpServers and load them with load_mcp_servers() + McpServerManager. See examples/07_mcp_tools.py for a complete, self-contained program.

See examples/01_basic.py for the smallest runnable program, and examples/06_host_and_permissions.py for a self-contained illustration of the host / permission model.

Environment variables

Variable	Purpose	Default
ANTHROPIC_API_KEY	Anthropic auth	—
OPENAI_API_KEY	OpenAI auth	—
CURRY_LEAVES_MODEL	Model id for the CLIs	auto-detected
CURRY_LEAVES_PROVIDER	Force a provider (anthropic | openai | ollama)	inferred from model id
CURRY_LEAVES_HOME	Base dir for settings / skills / sessions	~/.curry-leaves
CURRY_LEAVES_NO_RECORD	Disable session recording when set	recording on
OPENAI_BASE_URL	Point OpenAIProvider at any OpenAI-compatible gateway	https://api.openai.com/v1
OLLAMA_HOST	Ollama server URL	http://localhost:11434
NO_COLOR	Disable ANSI color	color on

Architecture

The design is a strict layering — a stateless definition on a stateful driver on a pure engine — with all I/O pushed to swappable seams (Provider, Host, Tool).

Layer	File	Responsibility
Message model	core/messages.py	Provider-neutral Message/Content types — the one thing everything agrees on.
Events	core/events.py	What the loop yields; a small structural set + a streaming delta payload.
Loop	core/loop.py	The pure engine: stream → run tools → stream, while tools are called. Yields events.
Tools	core/tools.py	A registry of pydantic-typed tools + a concurrent executor with a universal large-result guard.
Agent	core/agent.py	A stateless definition (model, tools, instructions, sub-agents, output_type).
Runner	runner.py	Live conversation state; builds the Context each turn; wires sub-agents, handoff, permissions, compaction.
Providers	providers/*	The only place that knows a wire format. Anthropic / OpenAI / Ollama behind one Provider.
MCP	mcp/*	MCP client: stdio/HTTP server connections, a manager, a settings loader, and a tool adapter.
Host	core/host.py	The frontend seam: emit(event) + request(req). Headless by default.
Prompt	prompt.py	Layered system prompt (identity → instructions → env → context → tools), cache-friendly.
Permission	permission.py	Per-call allow / ask / deny gate with standing approvals.
Thinking	thinking.py	A tiny classifier that sizes reasoning effort per task.
Skills	skills.py	Progressive-disclosure capability packages via skill:// refs.
Compaction	compaction.py	Summarizes old history as the context window fills.
Elision	elision.py	Stubs out stale tool results (recoverably) before compaction is needed.

The one decision that drives the whole loop:

runnable = stop_reason in ("tool_use", "stop") AND tool_calls exist

Tools were called → loop again. None → stop.

Project layout

src/curry_leaves/
  core/        # messages, events, loop, tools, agent, host, blobs — the kernel
  providers/   # anthropic, openai/ollama, factory, sse, base
  tools/       # read, write, edit, find, search, bash, tasks, ask, web, …
  mcp/         # MCP client: servers, manager, config, tool adapter
  session/     # session store + recording + forking (replay a transcript into a new branch)
  cli/         # chat REPL + Textual TUI
  util/        # paths, retry, frontmatter, resources
  runner.py prompt.py permission.py thinking.py skills.py compaction.py elision.py catalog.py settings.py
  __init__.py  # public API surface
examples/      # runnable examples
tests/         # pytest suite (MCP subsystem, session forking, elision)

Development

git clone https://github.com/Curry-Leaves/curry-leaves-py.git
cd curry-leaves-py
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"

mypy src               # strict type check (the correctness gate)
pytest                 # run the test suite
curry-leaves           # launch the Textual TUI (needs a TTY)
curry-leaves-repl      # launch the REPL (works with piped input)
python3 examples/01_basic.py "What is this project?"

mypy --strict is the gate every change must pass; pytest covers the MCP subsystem. Contributions extending the test suite are welcome.

Contributing

Contributions are very welcome — bug reports, features, docs, and tests. In short:

1. Open an issue first for anything non-trivial, so we can agree on the approach.
2. Fork & branch off main (git checkout -b feat/short-description).
3. Make the change, keep the diff focused, and ensure mypy src and pytest pass.
4. Open a pull request describing the what and why.

See CONTRIBUTING.md for the full guide — dev setup, code conventions, how to add a tool / provider / frontend capability, commit style, and bug-reporting.

Non-goals

curry-leaves is deliberately small. It does not include LSP integration, vector stores, or a plugin marketplace. If you need those, they belong in a layer built on top of the kernel, not inside it. (Unlike the TypeScript sibling, MCP is included here — as a thin client layer that feeds the existing Tool seam, not a change to the kernel.)

Acknowledgements

curry-leaves for Python is the sibling of curry-leaves-ts, a TypeScript kernel of the same design; the two ports mirror each other module-for-module.

License

MIT © Ilayanambi Ponramu