Langroid

Langroid follows an actor-style design: agents exchange messages, and each ChatAgent holds its own conversation state plus optional tools and a vector store. An agent has three responders, one each for the LLM, the agent's own tool handling, and the human user.

A Task wraps an agent with a system message and runs a loop over those responders until a done condition is met. Multi-agent systems are built by nesting: parent.add_sub_task(child) makes the child task another responder of the parent, tried in round-robin order after the agent's own responders. Because Task.run() has the same shape as a responder, delegation can be recursive, giving a tree of tasks rather than a flat group chat.

Memory is the agent's message history, with retrieval from vector stores (Qdrant, Chroma, LanceDB, Pinecone, PGVector, Weaviate) for document agents and optional Redis caching of LLM responses. Logs record message lineage so a reply can be traced to its origin.

The human is built into the loop as one of the three responders. By default a Task is interactive=True and waits for keyboard input after each non-human response. With interactive=False it runs unattended, except that an agent can still ask the person directly by addressing @user or using RecipientTool, and default_human_response can script replies for tests. No web-based approval UI is part of the core library.

pip install langroid

import langroid as lr
import langroid.language_models as lm

# Any OpenAI-compatible model works, e.g. chat_model="ollama/mistral" for a local one.
cfg = lr.ChatAgentConfig(llm=lm.OpenAIGPTConfig(chat_model=lm.OpenAIChatModel.GPT4o))

teacher = lr.Task(
    lr.ChatAgent(cfg), name="Teacher", interactive=False,
    system_message="Ask your student three short sums, one at a time. "
                   "After checking the third answer, reply DONE.",
)
student = lr.Task(
    lr.ChatAgent(cfg), name="Student", interactive=False, single_round=True,
    system_message="Answer each question with only the number.",
)
teacher.add_sub_task(student)
result = teacher.run(turns=20)

• The README asks for Python 3.11+; the package metadata allows >=3.10 and <3.14.
• Put OPENAI_API_KEY (or other provider settings) in a .env file; non-OpenAI models go through OpenAI-compatible servers or LiteLLM.
• Tasks default to interactive=True and wait for keyboard input after each agent reply; pass interactive=False for scripted runs.
• Features need extras: langroid[doc-chat], langroid[db], langroid[hf-embeddings] or langroid[all]. For SQLChatAgent on Postgres install langroid[postgres] (or psycopg2-binary if that fails).
• The README suggests reinstalling mysqlclient if you hit odd errors with it.

Yes, as a client. Its MCP adapter turns tools from MCP servers into Langroid ToolMessage classes. AG-UI works only through a community integration in the AG-UI repo, and no A2A support was found.

Yes. It is MIT licensed and has no paid tier; you pay only for the model API you use, or nothing when running local models.

Python only. The README asks for Python 3.11 or newer.

Each agent is wrapped in a Task, and tasks are nested with add_sub_task. A parent task treats its sub-tasks as extra responders and hands messages to them in turn, so delegation forms a tree.

The project publishes a Claude Code plugin (langroid@langroid) whose skills help Claude Code write Langroid code and record new patterns. Langroid itself does not drive Claude Code.