A unified agent orchestration hub that lets you configure and manage heterogeneous AI agents via YAML and expose them through standardized protocols.
You want to use multiple AI agents together - Claude Code for refactoring, Codex for code editing with advanced reasoning, a custom analysis agent, maybe Goose for specific tasks. But each has different APIs, protocols, and integration patterns. Coordinating them means writing glue code for each combination.
AgentPool acts as a protocol bridge. Define all your agents in one YAML file - whether they're native (PydanticAI-based), direct integrations (Claude Code, Codex), external ACP agents (Goose), or AG-UI agents. Then expose them all through ACP or AG-UI protocols, letting them cooperate, delegate, and communicate through a unified interface.
flowchart TB
subgraph AgentPool
subgraph config[YAML Configuration]
native[Native Agents<br/>PydanticAI]
direct[Direct Integrations<br/>Claude Code, Codex]
acp_agents[ACP Agents<br/>Goose, etc.]
agui_agents[AG-UI Agents]
workflows[Teams & Workflows]
end
subgraph interface[Unified Agent Interface]
delegation[Inter-agent delegation]
routing[Message routing]
context[Shared context]
end
config --> interface
end
interface --> acp_server[ACP Server]
interface --> opencode_server[OpenCode Server]
interface --> agui_server[AG-UI Server]
acp_server --> clients1[Zed, Toad, ACP Clients]
opencode_server --> clients2[OpenCode TUI/Desktop]
agui_server --> clients3[AG-UI Clients]
uv tool install agentpool
# agents.ymlagents:
assistant:
type: nativemodel: openai:gpt-4osystem_prompt: "You are a helpful assistant."# Run via CLI
agentpool run assistant "Hello!"# Or start as ACP server (for Zed, Toad, etc.)
agentpool serve-acp agents.ymlThe real power comes from mixing agent types:
agents:
# Native PydanticAI-based agentcoordinator:
type: nativemodel: openai:gpt-4otools:
- type: subagent # Can delegate to all other agentssystem_prompt: "Coordinate tasks between available agents."# Claude Code agent (direct integration)claude:
type: claude_codedescription: "Claude Code for complex refactoring"# Codex agent (direct integration)codex:
type: codexmodel: gpt-5.1-codex-maxreasoning_effort: mediumdescription: "Codex for code editing with advanced reasoning"# ACP protocol agentsgoose:
type: acpprovider: goosedescription: "Goose for file operations"# AG-UI protocol agentagui_agent:
type: aguiurl: "http://localhost:8000"description: "Custom AG-UI agent"Now coordinator can delegate work to any of these agents, and all are accessible through the same interface.
Agents can form teams (parallel) or chains (sequential):
teams:
review_pipeline:
mode: sequentialmembers: [analyzer, reviewer, formatter]parallel_coders:
mode: parallelmembers: [claude, goose]asyncwithAgentPool("agents.yml") aspool:
# Parallel executionteam=pool.get_agent("analyzer") &pool.get_agent("reviewer")
results=awaitteam.run("Review this code")
# Sequential pipelinechain=analyzer|reviewer|formatterresult=awaitchain.run("Process this")Everything is configurable - models, tools, connections, triggers, storage:
agents:
analyzer:
type: nativemodel:
type: fallbackmodels: [openai:gpt-4o, anthropic:claude-sonnet-4-0]tools:
- type: subagent
- type: resource_accessmcp_servers:
- "uvx mcp-server-filesystem"knowledge:
paths: ["docs/**/*.md"]connections:
- type: nodename: reporterfilter_condition:
type: word_matchwords: [error, warning]AgentPool can expose your agents through multiple server protocols:
| Server | Command | Use Case |
|---|---|---|
| ACP | agentpool serve-acp | IDE integration (Zed, Toad) - bidirectional communication with tool confirmations |
| OpenCode | agentpool serve-opencode | OpenCode TUI/Desktop - supports remote filesystems via fsspec |
| MCP | agentpool serve-mcp | Expose tools to other agents |
| AG-UI | agentpool serve-agui | AG-UI compatible frontends |
| OpenAI API | agentpool serve-api | Drop-in OpenAI API replacement |
The ACP server is ideal for IDE integration - it provides real-time tool confirmations and session management. The OpenCode server enables the OpenCode TUI to control AgentPool agents, including agents operating on remote environments (Docker, SSH, cloud sandboxes).
- Structured Output: Define response schemas inline or import Python types
- Storage & Analytics: Track all interactions with configurable providers
- File Abstraction: UPath-backed operations work on local and remote sources
- Triggers: React to file changes, webhooks, or custom events
- Streaming TTS: Voice output support for all agents
agentpool run agent_name "prompt"# Single run
agentpool serve-acp config.yml # ACP server for IDEs
agentpool serve-opencode config.yml # OpenCode TUI server
agentpool serve-mcp config.yml # MCP server
agentpool watch --config agents.yml # React to triggers
agentpool history stats --group-by model # View analyticsfromagentpoolimportAgentPoolasyncwithAgentPool("agents.yml") aspool:
agent=pool.get_agent("assistant")
# Simple runresult=awaitagent.run("Hello")
# Streamingasyncforeventinagent.run_stream("Tell me a story"):
print(event)
# Multi-modalresult=awaitagent.run("Describe this", Path("image.jpg"))For complete documentation including advanced configuration, connection patterns, and API reference, visit phil65.github.io/agentpool.