Framework

BeamAI · The AI agent framework for Erlang/OTP

BeamAI is the framework's core library, providing three foundations for building agents: ChatClient, a unified LLM client, and SimpleAgent. The design borrows from Semantic Kernel: ChatClient itself is stateless and records no messages — conversation history is managed by a Memory Filter keyed by conversation_id.

Erlang/OTP's lightweight processes and supervision trees make agents a fit for long-running, high-concurrency server workloads. Tool libraries, RAG, MCP, A2A, AG-UI and more live in BeamAI Extra.

ChatClient · Filter · LLM client · SimpleAgent

Erlang: create an LLM config and chat
LLM = beamai_chat_model:create(zhipu, #{
    model => <<"glm-4.7">>,
    api_key => list_to_binary(os:getenv("ZHIPU_API_KEY"))
}),

{ok, Response} = beamai_chat_model:chat(LLM, [
    {role, user, content, <<"Hello!"/utf8>>}
]).

Features

What it can do

  • One unified LLM client

    OpenAI, Anthropic, DeepSeek, Zhipu, DashScope, and Ollama behind a single sync and streaming interface; multimodal input, Anthropic caching / web search / citations, rate-limit headers, Retry-After-aware retries, and uniform error structures.

  • ChatClient and tools

    Semantic tool registration and invocation: declare name, description, parameters, and handler in one map; ChatClient hands them to the model, runs the call, and fills results back in.

  • Onion-style filters

    An interception chain around the model: insert logging, validation, rewriting, and safety checks before and after model and tool calls — independent and composable.

  • Conversation memory

    History is decoupled from ChatClient: send only the latest message each turn while a Memory Filter manages history by conversation_id. Storage is pluggable (ETS, persistent DETS, or your own behaviour); sliding windows are provided by the agent side.

  • SimpleAgent

    A ReAct-first agent: multi-turn dialogue, callback hooks, interruption and resume — a good fit for tool calls that need human confirmation.

  • Structured output

    The output parser turns model output into JSON, XML, or CSV, retrying automatically when parsing fails.

Specifications

Technical specifications

RuntimeErlang/OTP 27+, built with rebar3
Modulesbeamai_core (ChatClient, Filter, Tool), beamai_llm (LLM client), beamai_agent (SimpleAgent)
ExtensionsTools, RAG, MCP, A2A, AG-UI — see BeamAI Extra
LicenseApache License 2.0

FAQ

You might be wondering

BeamAI vs BeamAI Extra?

BeamAI is the core: ChatClient, the LLM client, and SimpleAgent. BeamAI Extra adds tool libraries, RAG, MCP, A2A, and AG-UI on top — pull in what you need.

Which model vendors?

OpenAI, Anthropic, DeepSeek, Zhipu, DashScope (Qwen), and Ollama built in, with identical sync and streaming interfaces. Other OpenAI-compatible services connect via base_url.

Where is conversation history stored?

Managed by the Memory Filter per conversation_id, with pluggable storage: ETS (memory), DETS (persistent), or your own behaviour.

What is the license?

Apache License 2.0, fine for commercial projects.

Open source

Apache 2.0 open source

The engine, tests, benchmarks, and design docs are all in the repository — read them, change them, file issues.

Related framework

BeamAI Extra · Tools, RAG, and protocol extensions for BeamAI

Extensions on the BeamAI core: file / shell / human-interaction tools and middleware, RAG, an MCP client and server, the A2A protocol, and an AG-UI event-stream server.