|
| 1 | +#Agent User Interaction (AG-UI) Protocol |
| 2 | + |
| 3 | +The[Agent User Interaction (AG-UI) Protocol](https://docs.ag-ui.com/introduction) |
| 4 | +is an open standard introduced by the |
| 5 | +[CopilotKit](https://webflow.copilotkit.ai/blog/introducing-ag-ui-the-protocol-where-agents-meet-users) |
| 6 | +team that standardises how frontend applications connect to AI agents through |
| 7 | +an open protocol. Think of it as a universal translator for AI-driven systems |
| 8 | +no matter what language an agent speaks: AG-UI ensures fluent communication. |
| 9 | + |
| 10 | +The team at[Rocket Science](https://www.rocketscience.gg/), contributed the |
| 11 | +initial version of AG-UI integration to make it easy to implement the AG-UI |
| 12 | +protocol with PydanticAI agents. |
| 13 | + |
| 14 | +This also includes an[`Agent.to_ag_ui`][pydantic_ai.Agent.to_ag_ui] convenience |
| 15 | +method which simplifies the creation of[`AGUIApp`][pydantic_ai.ag_ui.AGUIApp] |
| 16 | +for PydanticAI agents, which is built on top of[Starlette](https://www.starlette.io/), |
| 17 | +meaning it's fully compatible with any ASGI server. |
| 18 | + |
| 19 | +##Installation |
| 20 | + |
| 21 | +The only dependencies are: |
| 22 | + |
| 23 | +-[ag-ui-protocol](https://docs.ag-ui.com/introduction): to provide the AG-UI |
| 24 | + types and encoder. |
| 25 | +-[pydantic](https://pydantic.dev): to validate the request/response messages |
| 26 | +-[pydantic-ai](https://ai.pydantic.dev/): to provide the agent framework |
| 27 | + |
| 28 | +To run the examples you'll also need: |
| 29 | + |
| 30 | +-[uvicorn](https://www.uvicorn.org/) or another ASGI compatible server |
| 31 | + |
| 32 | +```bash |
| 33 | +pip/uv-add'uvicorn' |
| 34 | +``` |
| 35 | + |
| 36 | +You can install PydanticAI with the`ag-ui` extra to ensure you have all the |
| 37 | +required AG-UI dependencies: |
| 38 | + |
| 39 | +```bash |
| 40 | +pip/uv-add'pydantic-ai-slim[ag-ui]' |
| 41 | +``` |
| 42 | + |
| 43 | +##Quick start |
| 44 | + |
| 45 | +```py {title="agent_to_ag_ui.py" py="3.10" hl_lines="17-28"} |
| 46 | +"""Basic example for AG-UI with FastAPI and Pydantic AI.""" |
| 47 | + |
| 48 | +from__future__import annotations |
| 49 | + |
| 50 | +from pydantic_aiimport Agent |
| 51 | + |
| 52 | +agent= Agent('openai:gpt-4.1',instructions='Be fun!') |
| 53 | +app= agent.to_ag_ui() |
| 54 | +``` |
| 55 | + |
| 56 | +You can run the example with: |
| 57 | + |
| 58 | +```shell |
| 59 | +uvicorn agent_to_ag_ui:app --host 0.0.0.0 --port 8000 |
| 60 | +``` |
| 61 | + |
| 62 | +This will expose the agent as an AG-UI server, and you can start sending |
| 63 | +requests to it. |
| 64 | + |
| 65 | +##Design |
| 66 | + |
| 67 | +The solution provides and adapter layer between the AG-UI protocol and |
| 68 | +PydanticAI agents written in Python, including support for all aspects of spec |
| 69 | +including: |
| 70 | + |
| 71 | +-[Events](https://docs.ag-ui.com/concepts/events) |
| 72 | +-[Messages](https://docs.ag-ui.com/concepts/messages) |
| 73 | +-[State Management](https://docs.ag-ui.com/concepts/state) |
| 74 | +-[Tools](https://docs.ag-ui.com/concepts/tools) |
| 75 | + |
| 76 | +The adapter receives messages in the form of a |
| 77 | +[`RunAgentInput`](https://docs.ag-ui.com/sdk/js/core/types#runagentinput) |
| 78 | +which describes the details of a request being passed to the agent including |
| 79 | +messages and state. These are then converted to PydanticAI types, passed to the |
| 80 | +agent which then process the request. |
| 81 | + |
| 82 | +Results from the agent are converted from PydanticAI types to AG-UI events and |
| 83 | +streamed back to the caller as Server-Sent Events (SSE). |
| 84 | + |
| 85 | +A user request may require multiple round trips between client UI and PydanticAI |
| 86 | +server, depending on the tools and events needed. |
| 87 | + |
| 88 | +This is exposed via the[AGUIApp][pydantic_ai.ag_ui.AGUIApp] which is slim |
| 89 | +wrapper around[Starlette](https://www.starlette.io/) providing easy access to |
| 90 | +run a PydanticAI server with AG-UI support with any ASGI server. |
| 91 | + |
| 92 | +##Features |
| 93 | + |
| 94 | +To expose a PydanticAI agent as an AG-UI server including state support, you |
| 95 | +use the[`to_ag_ui`][pydantic_ai.agent.Agent.to_ag_ui] method create an ASGI |
| 96 | +compatible server. |
| 97 | + |
| 98 | +In the example below we have document state which is shared between the UI and |
| 99 | +server using the[`StateDeps`][pydantic_ai.ag_ui.StateDeps] which implements the |
| 100 | +[`StateHandler`][pydantic_ai.ag_ui.StateHandler] that can be used to automatically |
| 101 | +decode state contained in[`RunAgentInput.state`](https://docs.ag-ui.com/sdk/js/core/types#runagentinput) |
| 102 | +when processing requests. |
| 103 | + |
| 104 | +###State management |
| 105 | + |
| 106 | +The adapter provides full support for |
| 107 | +[AG-UI state management](https://docs.ag-ui.com/concepts/state), which enables |
| 108 | +real-time synchronization between agents and frontend applications. |
| 109 | + |
| 110 | +```python {title="ag_ui_state.py" py="3.10" hl_lines="18-40"} |
| 111 | +"""State example for AG-UI with FastAPI and Pydantic AI.""" |
| 112 | + |
| 113 | +from__future__import annotations |
| 114 | + |
| 115 | +from pydanticimport BaseModel |
| 116 | + |
| 117 | +from pydantic_aiimport Agent |
| 118 | +from pydantic_ai.ag_uiimport StateDeps |
| 119 | + |
| 120 | + |
| 121 | +classDocumentState(BaseModel): |
| 122 | +"""State for the document being written.""" |
| 123 | + |
| 124 | + document:str='' |
| 125 | + |
| 126 | + |
| 127 | +agent= Agent( |
| 128 | +'openai:gpt-4.1', |
| 129 | +instructions='Be fun!', |
| 130 | +deps_type=StateDeps[DocumentState], |
| 131 | +) |
| 132 | +app= agent.to_ag_ui(deps=StateDeps(DocumentState())) |
| 133 | +``` |
| 134 | + |
| 135 | +Since`app` is an ASGI application, it can be used with any ASGI server e.g. |
| 136 | + |
| 137 | +```bash |
| 138 | +uvicorn agent_to_ag_ui:app --host 0.0.0.0 --port 8000 |
| 139 | +``` |
| 140 | + |
| 141 | +Since the goal of[`to_ag_ui`][pydantic_ai.agent.Agent.to_ag_ui] is to be a |
| 142 | +convenience method, it accepts the same a combination of the arguments require |
| 143 | +for: |
| 144 | + |
| 145 | +-[`AGUIApp`][pydantic_ai.ag_ui.AGUIApp] constructor |
| 146 | +-[`Agent.iter`][pydantic_ai.agent.Agent.iter] method |
| 147 | + |
| 148 | +###Tools |
| 149 | + |
| 150 | +AG-UI tools are seamlessly provided to the PydanticAI agent, enabling rich |
| 151 | +use experiences with frontend user interfaces. |
| 152 | + |
| 153 | +###Events |
| 154 | + |
| 155 | +The adapter provides the ability for PydanticAI tools to send |
| 156 | +[AG-UI events](https://docs.ag-ui.com/concepts/events) simply by defining a tool |
| 157 | +which returns a type based off |
| 158 | +[`BaseEvent`](https://docs.ag-ui.com/sdk/js/core/events#baseevent) this allows |
| 159 | +for custom events and state updates. |
| 160 | + |
| 161 | +```python {title="ag_ui_tool_events.py" py="3.10" hl_lines="34-55"} |
| 162 | +"""Tool events example for AG-UI with FastAPI and Pydantic AI.""" |
| 163 | + |
| 164 | +from__future__import annotations |
| 165 | + |
| 166 | +from ag_ui.coreimport CustomEvent, EventType, StateSnapshotEvent |
| 167 | +from pydanticimport BaseModel |
| 168 | + |
| 169 | +from pydantic_aiimport Agent, RunContext |
| 170 | +from pydantic_ai.ag_uiimport StateDeps |
| 171 | + |
| 172 | + |
| 173 | +classDocumentState(BaseModel): |
| 174 | +"""State for the document being written.""" |
| 175 | + |
| 176 | + document:str='' |
| 177 | + |
| 178 | + |
| 179 | +agent= Agent( |
| 180 | +'openai:gpt-4.1', |
| 181 | +instructions='Be fun!', |
| 182 | +deps_type=StateDeps[DocumentState], |
| 183 | +) |
| 184 | +app= agent.to_ag_ui(deps=StateDeps(DocumentState())) |
| 185 | + |
| 186 | + |
| 187 | +@agent.tool |
| 188 | +defupdate_state(ctx: RunContext[StateDeps[DocumentState]]) -> StateSnapshotEvent: |
| 189 | +return StateSnapshotEvent( |
| 190 | +type=EventType.STATE_SNAPSHOT, |
| 191 | +snapshot=ctx.deps.state, |
| 192 | + ) |
| 193 | + |
| 194 | + |
| 195 | +@agent.tool_plain |
| 196 | +defcustom_events() -> list[CustomEvent]: |
| 197 | +return [ |
| 198 | + CustomEvent( |
| 199 | +type=EventType.CUSTOM, |
| 200 | +name='count', |
| 201 | +value=1, |
| 202 | + ), |
| 203 | + CustomEvent( |
| 204 | +type=EventType.CUSTOM, |
| 205 | +name='count', |
| 206 | +value=2, |
| 207 | + ), |
| 208 | + ] |
| 209 | +``` |
| 210 | + |
| 211 | +##Examples |
| 212 | + |
| 213 | +For more examples of how to use[`to_ag_ui`][pydantic_ai.Agent.to_ag_ui] see |
| 214 | +[`pydantic_ai_ag_ui_examples`](https://github.com/pydantic/pydantic-ai/tree/main/examples/pydantic_ai_ag_ui_examples), |
| 215 | +which includes working server for the with the |
| 216 | +[AG-UI Dojo](https://docs.ag-ui.com/tutorials/debugging#the-ag-ui-dojo) which |
| 217 | +can be run from a clone of the repo or with the`pydantic-ai-examples` package |
| 218 | +installed with either of the following: |
| 219 | + |
| 220 | +```bash |
| 221 | +pip/uv-add pydantic-ai-examples |
| 222 | +``` |
| 223 | + |
| 224 | +Direct, which supports command line flags: |
| 225 | + |
| 226 | +```shell |
| 227 | +python -m pydantic_ai_ag_ui_examples.dojo_server --help |
| 228 | +usage: dojo_server.py [-h] [--port PORT] [--reload] [--no-reload] [--log-level {critical,error,warning,info,debug,trace}] |
| 229 | + |
| 230 | +PydanticAI AG-UI Dojo server |
| 231 | + |
| 232 | +options: |
| 233 | + -h, --help show thishelp message andexit |
| 234 | + --port PORT, -p PORT Port to run the server on (default: 9000) |
| 235 | + --reload Enable auto-reload (default: True) |
| 236 | + --no-reload Disable auto-reload |
| 237 | + --log-level {critical,error,warning,info,debug,trace} |
| 238 | + Agent log level (default: info) |
| 239 | +``` |
| 240 | + |
| 241 | +Run with dojo server with debug logging: |
| 242 | + |
| 243 | +```shell |
| 244 | +python -m pydantic_ai_ag_ui_examples.dojo_server --log-level debug |
| 245 | +``` |
| 246 | + |
| 247 | +Using uvicorn: |
| 248 | + |
| 249 | +```shell |
| 250 | +uvicorn pydantic_ai_ag_ui_examples.dojo_server:app --port 9000 |
| 251 | +``` |
| 252 | + |
| 253 | +There is also a simplified basic example: |
| 254 | + |
| 255 | +```shell |
| 256 | +python -m pydantic_ai_ag_ui_examples.basic |
| 257 | +``` |