microagents — minimal, modular AI-agent framework in Rust

Features

🔌 Multi-provider LLM support

OpenAI, OpenRouter, Groq, and Ollama out of the box via ultrafast-models-sdk.

🛠 Tool-use loop

Agents call tools, wait for results, and continue the conversation. Parallel tool calls supported.

📦 Skills

Drop-in markdown skill packs with front-matter loaded from ./.agents/skills or ~/.agents/skills.

💾 Session storage

Resume conversations with in-memory, JSONL, or SQLite backends.

🔍 Hybrid search

Semantic + sparse (BM25) search over your workspace, powered by Qdrant, fastembed, and astchunk.

🖥 Embedded TUI

Ratatui-based chat interface with streaming deltas, tool-call visualization, and session resume.

Architecture

microagents-cli

binary: microag

TUI + built-in tools + vector search indexing

microagents-core

Agent trait, MicroAgent builder, tool loop, skill loader

microagents-events

Structured events → JSON-RPC notifications

microagents-storage

In-memory / JSONL / SQLite session backends

Quick Start

Run the TUI with default settings, or customize provider, model, and storage:

# Default TUI
microag

# Custom provider & model
microag --provider ollama --model gemma4:latest --storage sqlite

# Resume a session
microag --session-id <session-id>

# Headless mode (JSON-RPC output)
microag -p 'Where is AgentStorage defined?'

# Verbose indexing
microag --verbose -p 'Which storage backends implement AgentStorage?'

Programmatic Usage

use microagents_core::{
    agent::{MicroAgentBuilder, SupportedProvider},
    types::{Agent, ToolExecutionContext},
};
use microagents_storage::types::AgentStorageChoice;
use std::sync::Arc;

let agent = MicroAgentBuilder::new(ToolExecutionContext::new(()))
    .provider("openrouter".to_string())?
    .model("anthropic/claude-opus-4.7".to_string())
    .storage(AgentStorageChoice::Sqlite)
    .await?
    .find_skills()?
    .build()?;

let mut stream = agent.run("Hello!".to_string(), None).await?;
while let Some(event) = stream.next().await { /* … */ }

Workspace Crates

Crate	Purpose
`microagents-core`	Agent definition, tool trait, skill loading, LLM generation loop
`microagents-events`	Event types (`AgentEventAny`, `SessionInitEvent`, …) and JSON-RPC serialization
`microagents-storage`	Session persistence: in-memory, JSONL, or SQLite
`microagents-cli`	Interactive terminal UI + built-in tools (`read`, `write`, `edit`, `shell_execute`, `search`)