codex-python

Python SDK for Codex with bundled codex binaries inside platform wheels.

This package exposes two supported APIs:

• Codex: a simple, local convenience interface backed by a private stdio app-server session
• AppServerClient: a richer app-server client for thread management, streaming events, approvals, and typed protocol access

Canonical import paths:

• use from codex import ... for the high-level Codex facade
• use from codex.app_server import ... for the raw app-server client and app-server option types

Install

pip install codex-python

Which API should I use?

Codex

Use Codex when you want the smallest surface area for local automation:

• one private local app-server session per Codex instance
• stateless run*() convenience (fresh internal thread per call)
• stateful thread workflows when needed via start_thread() / resume_thread()
• simple request/response usage
• optional streaming over the exec event stream
• structured output via TurnOptions(output_schema=...)

AppServerClient

Use AppServerClient when you want a deeper integration:

• persistent app-server connection
• thread objects and turn streams
• protocol-native notifications
• server-driven requests such as tool callbacks and approvals
• typed protocol models and raw JSON-RPC access when needed

Quickstart: Codex

from codex import Codex, ThreadStartOptions

client = Codex()

# Simplest one-shot call.
summary = client.run_text("Diagnose the failing tests and propose a fix")
print(summary)

# One-shot call with thread-scoped defaults for that run's fresh internal thread.
summary = client.run_text(
    "Diagnose the failing tests in this repo",
    thread_options=ThreadStartOptions(
        cwd="/repo",
        model="gpt-5",
    ),
)
print(summary)

Use thread_options= on run(), run_text(), run_json(), and run_model() when you want to set defaults on the fresh internal thread created for that one-shot call. Use start_thread() / resume_thread() when later runs should share context.

More Codex examples: docs/exec_api.md

Quickstart: AppServerClient

from codex.app_server import AppServerClient, AppServerClientInfo, AppServerInitializeOptions

initialize_options = AppServerInitializeOptions(
    client_info=AppServerClientInfo(
        name="my_integration",
        title="My Integration",
        version="0.1.0",
    )
)

with AppServerClient.connect_stdio(initialize_options=initialize_options) as client:
    thread = client.start_thread()
    summary = thread.run_text("Briefly summarize this repository's purpose.")
    print(summary)

More app-server examples: docs/app_server.md For websocket transport, install the optional extra: pip install "codex-python[websocket]".

Dynamic tools

Decorator-driven dynamic tools are available on both SDK surfaces.

Codex

from codex import Codex, dynamic_tool


@dynamic_tool
def lookup_ticket(id: str) -> str:
    """Look up a support ticket by id."""
    return f"Ticket {id}: Login requests time out in eu-west-1."


client = Codex()
summary = client.run_text(
    "Use the lookup_ticket dynamic tool for ticket 123 and summarize the result.",
    tools=[lookup_ticket],
)
print(summary)

AppServerClient

For a complete app-server example using the same decorator-driven flow, see examples/app_server_dynamic_tool.py.

Structured output

Codex

from codex import Codex, TurnOptions

schema = {
    "type": "object",
    "properties": {"summary": {"type": "string"}},
    "required": ["summary"],
    "additionalProperties": False,
}

client = Codex()
payload = client.run_json("Summarize repository status", TurnOptions(output_schema=schema))
print(payload["summary"])

AppServerClient

from pydantic import BaseModel

from codex.app_server import AppServerClient, AppServerTurnOptions


class Summary(BaseModel):
    summary: str


with AppServerClient.connect_stdio() as client:
    thread = client.start_thread()
    result = thread.run_model(
        "Summarize repository status",
        Summary,
    )
    print(result.summary)

run_model() uses Summary both as the validation model and, by default, as the output schema sent to Codex. If you want JSON back without validation, you can also pass the model class directly to output_schema, for example thread.run_json(..., AppServerTurnOptions(output_schema=Summary)).

Streaming

Codex stream

from codex import Codex
from codex.protocol import types as protocol

client = Codex()
stream = client.run("Investigate this bug")
for event in stream:
    if isinstance(event, protocol.ItemAgentMessageDeltaNotification):
        print(event.params.delta, end="", flush=True)

print()

Codex.run*() starts a fresh internal thread for each call. Use start_thread() or resume_thread() when you want later runs to share context.

High-level Codex helpers raise ThreadRunError on failed or interrupted terminal turns and preserve the final turn metadata on the exception for debugging and UI handling.

App-server stream

from codex.app_server import AppServerClient
from codex.protocol import types as protocol

with AppServerClient.connect_stdio() as client:
    thread = client.start_thread()
    stream = thread.run("Investigate this bug")

    for event in stream:
        if isinstance(event, protocol.ItemAgentMessageDeltaNotification):
            print(event.params.delta, end="", flush=True)

    print()

Advanced app-server usage, including typed stable RPC domains such as client.models and the raw client.rpc fallback: docs/app_server_advanced.md

Examples

• examples/basic_conversation.py: minimal Codex flow
• examples/basic_dynamic_tool.py: high-level Codex dynamic tool flow
• examples/app_server_conversation.py: minimal app-server flow
• examples/app_server_websocket_conversation.py: minimal websocket app-server flow
• examples/app_server_stream_events.py: protocol-native app-server streaming
• examples/app_server_tool_handler.py: typed app-server request handling
• examples/app_server_dynamic_tool.py: decorator-driven dynamic tool registration

Bundled binary behavior

By default, the SDK resolves the bundled binary at:

codex/vendor/<target-triple>/codex-app-server/{codex-app-server|codex-app-server.exe}

The bundled app-server runs directly. If it is not present, for example in a source checkout, the SDK falls back to codex app-server using codex on PATH.

You can override the executable path with:

• CodexOptions(codex_path_override=...)
• codex.app_server.AppServerProcessOptions(codex_path_override=...)

Overrides named codex-app-server* run directly; other overrides run with the app-server subcommand.

Development

make lint
make test

make test emits a terminal coverage report, writes coverage.xml, and enforces the repository coverage gate.

If you want to test vendored-binary behavior locally, fetch binaries into codex/vendor:

python scripts/fetch_codex_binary.py --target-triple x86_64-unknown-linux-musl