Not another AI Agent, but an open application framework for building Agents — purpose-built for embedded, edge, and vertical industries, equally suited for highly customizable, high-performance cloud and desktop AI applications. Simple, pure, predictable.

Unlike LangChain, CrewAI, or AutoGen, phi-agent ships with zero built-in tools. No pre-packaged toolkits, no hidden prompt engineering, no magic workflow engine — just a clean Rust runtime. You define every tool, you control every behavior.

Built on agent-base and agent-works. phi-agent provides the infrastructure. You bring the tools.

Ecosystem

phi-agent is part of a family of independent crates:

Crate	crates.io	Description
agent-base		Lightweight runtime kernel — LLM clients, Tool trait, event stream
agent-works		Batteries-included toolbox — MCP, Skills, Focus
phi-agent		Full framework — builder factory, renderers, config, CLI binary

Just need the runtime? cargo add agent-base. Need the full framework? cargo add phi-agent.

Architecture

graph TB
    AB[agent-base<br/>Tool trait · Runtime<br/>LLM clients · Events]

    AB --> AW[agent-works<br/>MCP · Skills · Focus]
    AB --> PT[phi-tools<br/>LocalShellTool]
    AB --> YT[your-tools<br/>Custom Tool impls]

    AW --> PA[phi-agent<br/>Builder factory<br/>Renderers · Config · Session<br/>CLI binary]

    PT --> PA
    YT --> PA

    PA --> Terminal[Terminal REPL]
    PA --> JSON[JSON Stream]
    PA --> Web[Web Backend]

Core principle: phi-agent ships with zero built-in tools. You define them, you register them. phi-agent discovers and manages them at runtime — listing, logging, and routing tool calls automatically.

Why phi-agent

Built for Vertical Scenarios. Not a generic chatbot, but an Agent framework for embedded systems, industrial, IoT, and other vertical domains, as well as desktop and cloud applications that demand deep customization — your scenario, your tools, your full control.

Lightweight, Runs Anywhere. A single Rust binary with zero runtime dependencies — from embedded Linux and edge gateways to cloud containers and desktop applications, cargo install gets you started in seconds, deploy anywhere.

Zero Built-in Tools, Fully Customizable. No pre-packaged tools, no platform lock-in — a tool is just 3 methods: name(), definition(), call(), you register what you need, the Agent uses what you register, only bring what your scenario truly needs, LLM freedom, precise and clean.

Fully Observable, Every Step Explainable. Every decision is logged, every step is traceable, with built-in session logging, structured tracing, and session metrics at a glance — compliance and audit trails without the stress.

Features

• Builder factory — base_agent_builder() with sensible defaults (thinking, recovery, limits)
• Three renderers — Terminal (rich, colored, streaming), JSON stream (JSONL), Null (silent)
• CLI-ready — REPL and one-shot modes with 20+ configurable flags
• Session management — auto-cleanup, file locking, JSONL turn logging
• Tool-agnostic — no built-in tools; register your own via AgentBuilder
• Extensible — middleware, approval handlers, custom renderers

Quick Start

use phi_agent::{
    base_agent_builder, build_system_prompt, PhiAgent, PhiAgentConfig,
    OpenAiClient, SafetyConfig, ReasoningEffort,
};
use std::sync::Arc;

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    // 1. Create LLM client
    let llm_client = Arc::new(OpenAiClient::new(
        std::env::var("LLM_API_KEY")?,
        "opus".into(),
        Some("https://api.openai.com/v1".into()),
    ));

    // 2. Build agent (register your tools here)
    let builder = base_agent_builder(llm_client)
        .system_prompt(build_system_prompt())
        .register_tool(your_tool);

    let agent = PhiAgent::build(builder, PhiAgentConfig {
        model: "opus".into(),
        enable_thinking: true,
        thinking_budget: None,
        thinking_effort: ReasoningEffort::Medium,
        safety: SafetyConfig::default(),
    })?;

    // 3. Run
    let session = agent.create_session().await;
    let renderer = phi_agent::create_stdout_renderer(
        &phi_agent::OutputFormat::Terminal {
            show_thinking: true,
            show_tool_args: true,
            color: true,
        }
    );

    agent.run_turn(session, "Hello!", |event| {
        renderer.render(event)
    }).await?;

    Ok(())
}

See examples/ for more complete examples.

CLI

cargo install phi-agent
phi "What's in this directory?"

# REPL mode
phi

# JSON output for scripting
phi --format json "list files"

Custom Tool Example

use agent_base::{Tool, ToolContext, ToolOutput, ToolControlFlow, AgentResult};
use serde_json::{Value, json};
use async_trait::async_trait;

struct HelloTool;

#[async_trait]
impl Tool for HelloTool {
    fn name(&self) -> &'static str { "hello" }

    fn definition(&self) -> Value {
        json!({
            "name": "hello",
            "description": "Say hello to someone",
            "parameters": {
                "type": "object",
                "properties": {
                    "name": { "type": "string", "description": "Who to greet" }
                },
                "required": ["name"]
            }
        })
    }

    async fn call(&self, args: &Value, _ctx: &ToolContext) -> AgentResult<ToolOutput> {
        let name = args["name"].as_str().unwrap_or("world");
        Ok(ToolOutput { summary: format!("Hello, {}!", name), control_flow: ToolControlFlow::Continue, raw: None, truncation: None })
    }
}

Full guide: guide/custom-tool.md

Documentation

📖 Full documentation: docs.phi-agent.dev

FAQ

Q: What's the difference between phi-agent and agent-base?

agent-base is the runtime kernel (LLM calls, tool orchestration, event stream). phi-agent wraps it with a builder factory, renderers, config resolution, and session management — plus a CLI binary.

Q: Can I use phi-agent without the CLI?

Yes. Import it as a library (phi_agent) and use PhiAgent::build() programmatically. The CLI is just one consumer.

Q: How do I add my own tools?

Implement the Tool trait from agent-base and register with builder.register_tool(...). phi-agent has zero knowledge of what tools exist.

Q: Does phi-agent support Anthropic / other providers?

Yes. agent-base provides AnthropicClient and OpenAiClient. Any client implementing LlmClient works.

Contributing

git clone git@github.com:hibuka-labs/phi-agent.git
cd phi-agent
cargo check

See CONTRIBUTING.md for detailed setup instructions and PR guidelines.

Contributors

Thanks goes to these wonderful people:

(emoji key) — This project follows the all-contributors specification.

License

MIT — see LICENSE for details.

Contact

:material-email-outline: phiagent@hibuka.com

中文文档