pycodex
pycodex brings the Codex agent loop to Python, with a CLI, reusable Agent API and browser workspace. It targets upstream default behavior for the selected features, model and protocol; supported features and deliberate differences are documented in Alignment.
Intentional boundaries include no sandbox enforcement or collaboration-mode switching, no OpenAI-specific model/effort allowlists for sub-agents, and no requirement to mirror every OpenAI model-catalog entry.
| Package | Provides | Python |
|---|---|---|
python-codex |
Agent, runtime, CLI, tools, Responses compatibility server and portable storage | 3.6.2+ |
pycodex-ws |
Browser workspace, Canvas, file/diff viewers and optional Feishu integration | 3.7+ |
The workspace package installs the matching core version. Python imports use
pycodex; the commands are pycodex and pycodex-ws.
The exec tool runs Python with top-level await in the host process. Each cell
has its own namespace, output and tracked asyncio tasks; imports and process
state are shared. Stopping is cooperative, so run blocking work through
tools.exec_command. Direct file access can use the cell's cwd Path.
See exec lifecycle and limits.
Install and run
Install the core CLI:
python3 -m pip install python-codex
pycodex
Or install the browser workspace, which includes core:
python3 -m pip install pycodex-ws
pycodex-ws --help
Use python3 -m pip install --upgrade pycodex-ws to upgrade both packages
together. If a command is missing from PATH, use python3 -m pycodex or
python3 -m workspace_server with the interpreter used for installation.
Model configuration
pycodex reads ~/.codex/config.toml and the adjacent .env. Configure model,
model_provider, and the provider's base_url and env_key as needed.
The built-in openai provider uses OPENAI_API_KEY. Authentication uses API
keys; ChatGPT sign-in credentials from auth.json or the OS keychain are not
loaded.
--profile NAME selects an adjacent NAME.config.toml or an embedded
[profiles.NAME] table. Both recursively merge over the base config, including
provider and feature tables. A top-level profile = "NAME" sets the default;
--profile overrides it. Duplicate names across both formats and missing selected
profiles are errors. Existing embedded profiles can be used without migration.
pycodex --profile my_profile "Summarize this repository."
pycodex --config /path/to/config.toml
pycodex doctor
Models use the streaming Responses API; OpenAI models use native Responses
directly. For a vLLM Chat Completions backend, set use_chat_completion = true
in its provider table. Other backend options include:
pycodex --vllm-endpoint http://127.0.0.1:18000
pycodex --profile my_profile --use-messages
See context and configuration and the Responses compatibility server. The compatibility server supports namespaced function/custom tools, including upstream Codex subagent tools, with namespace restoration on returned calls. Text and reasoning streams include item lifecycle events for Codex's live display.
For step-5-preview on Step's step_plan/v1 endpoint, enable
use_chat_completion = true in the Step provider or profile: its native Responses
endpoint does not expose namespaced collaboration tools correctly. The adapter
presents names such as collaboration__spawn_agent as ordinary functions.
Prompts follow the declared direct-call names and describe Python exec
collaboration only when both exec and collaboration tools are visible.
CLI
Run pycodex without a prompt for an interactive session, or supply a prompt
for a single turn:
pycodex "Summarize this repository in one sentence."
printf 'Reply with exactly OK.' | pycodex
pycodex --json "Reply with exactly OK."
| Command | Purpose |
|---|---|
/model, /effort |
Show or change the model and reasoning effort |
/cd <directory> |
Change the session's working directory while idle |
/title, /history |
Manage the title and inspect conversation history |
/resume |
List saved sessions; pass an index or session UUID to resume |
/compact |
Summarize history to free context |
/queue <message> |
Wait for the active turn to finish before submitting |
Ordinary input during a turn steers at the next safe sampling boundary.
Slash input uses Codex's command rules: /data/project and /tmp/ are
ordinary messages, while an unknown single-component name such as /modle
or /tmp reports a command error. Start the input with a space to send any
slash-prefixed text literally, including known commands such as /model.
These rules also apply to the browser workspace and Feishu.
Sessions are saved under CODEX_HOME/sessions. Set
model_auto_compact_token_limit in the configuration to compact automatically.
Ctrl+D or Ctrl+C stops the session and its child agents: pending turns are
discarded, and active turns stop after issued model/tool operations finish.
Session close uses the same behavior. A second Ctrl+C during closing exits
immediately. Use pycodex --help for startup options.
Browser workspace
Create workspaces.json:
{
"workspaces": [
{"id": "demo", "work_dir": ".", "state_path": "demo", "enable_board": true}
]
}
pycodex-ws --listen 127.0.0.1:6007 --workspace-config workspaces.json
Open http://127.0.0.1:6007. The workspace combines conversations with a live
HTML Canvas and read-only file/diff viewers. Click file links in replies to open
them, and use Insert into chat to add selected code to an editable draft.
Text and image views support files up to 10 MiB each.
The context ring opens the built-in Agent tree with n running · m failed counts
(including the root), each Agent's turn duration, and error details for failures.
Click a root or child Agent in the popover to open its live transcript tab.
Compare local files with diff:left..right; either path can be absolute or relative.
Use [TODO](TODO.md) for a local file link; bare filenames stay as text.
Local PDF links open in the browser's built-in reader, with Open in new tab
and Download links. PDF files are streamed separately from text snapshots.
Keep the server's stdin open: EOF initiates shutdown.
New session uses the launch --profile, then the config's default profile,
then the base config. The + button on its right (Custom profile) lets
you choose from adjacent *.config.toml files and embedded [profiles.NAME]
tables; Default uses the same default selection.
Each session saves its profile name and uses it again on restart; existing sessions
keep their configuration. Restart the server after adding or removing profiles.
The page loads the last selected session directly on reopening. Switching back
to a visited session immediately restores its rendered messages and scroll
position, then refreshes its content from the server.
While the selected Agent is working, the empty composer cycles through usage
hints about commands, frontend customization, workspace sharing, Agent status,
live transcripts and optional features, including clock tick dividers, about every
six seconds, starting at a random hint. Edit the short
strings in the hints list in workspace.html's updatePromptPlaceholder to change the tips.
The gear button opens Customize Agent and Share, with a short description of each.
Customize Agent starts an Agent conversation for editing files in
~/.pycodex/frontend/. Its setup context includes an API quick
reference and the path to the full API documentation shipped with the package.
Browse the feature catalog
in the view area and tell the Agent which features to adapt. You can also ask it
to publish a local customization as a GitHub feature PR using the
contribution format.
The custom frontend API provides structured state,
command results, history pagination, event replay and Agent/background-task
inventories. Optional custom.js extensions can export mount(host) to use
the built-in page's session, draft and viewer controls.
With a workspace server administrator password configured, Share sets a
separate password for this workspace using the existing login page. Sharing
passwords and visitor logins live only in memory; changing the password, stopping
sharing or restarting the process invalidates them. Router login is unchanged.
See the workspace guide for session controls, storage and Feishu setup. The separately distributed VS Code extension uses the same workspace backend, including on Remote SSH hosts.
Python API
Use the standard model and tool setup from Python:
import atexit
from pycodex import AgentsManager, CodexConfig
from pycodex.bootstrap import build_model
from pycodex.utils.async_bridge import run_async
async def main():
config = CodexConfig.from_env()
bootstrap = build_model(config.provider)
if bootstrap.owned_server is not None:
atexit.register(bootstrap.owned_server.stop)
runtime = AgentsManager(
bootstrap.client, config.context,
).create_root_runtime()
await runtime.start()
try:
receipt = await runtime.submit_input("Summarize this repository in one sentence.")
print((await receipt.future).output_text)
finally:
await runtime.close()
run_async(main())
CodexConfig.from_env(path, profile) loads the selected TOML and adjacent .env
once for startup. Pass config.provider to build_model and config.context to
AgentsManager; use dataclasses.replace for explicit overrides, including None.
API keys are still read from the environment when requests are sent.
build_model returns ModelBootstrapResult(client, owned_server). The caller owns
any locally started compatibility server; close() stops only that server.
CLI and workspace entry points retain process-exit cleanup for these services.
For custom integrations, Agent requires a model client, ToolRegistry,
ContextConfig and session_id. await agent.run_turn([...]) runs an asynchronous
turn; AgentRuntime(agent, agents_manager) adds input queues, commands and frontend
subscriptions. AgentsManager owns the tree of named tasks and supplies the
standard construction path.
The current prototype uses MultiAgentV2 tools: spawn_agent, send_message,
followup_task, wait_agent, interrupt_agent and list_agents. Child context
inheritance uses fork_turns; the session /fork command has been removed.
features.multi_agent_v2.max_concurrent_threads_per_session sets the total
capacity, defaulting to 16 including the root. When full, spawn closes the least
recently used idle leaf with no pending or background work, or reports a capacity
error. Closed children are removed from the Agent tree and Workspace views;
their task names can be reused. The Manager retains closing child references
until their cleanup finishes, even after they leave the live tree.
Recorded sessions retain the clock period in a small environment snapshot beside
the rollout. Resuming starts a fresh countdown and restores idle: sleeping
in attached frontends; child Agents remain in memory.
Runtime reads execution facts through Agent.execution_snapshot(); session
history and live-state snapshots retain their separate frontend formats.
Tool factories register shared resource managers with the environment, which
owns their callbacks, restart, background state and cleanup. Tools provide their
model-callable behavior. Pass tools_factory(manager, environment, toolset) to
AgentsManager to construct custom tools for both root and child Agents; see the
resource contract.
Model clients expose explicit recording metadata; Agent supplies each request's
session identity through Prompt, including compaction and resumed sessions.
Tool summaries and conversation-history projections are pure helpers shared by
event and session views.
See runtime contracts, context assembly and
the versioned alignment baseline and remaining differences.
Portable configuration
Portable mode transfers an encrypted bundle of your configuration, instructions and skills to another machine. Start a storage service in a separate terminal:
python3 -m pycodex.portable.server --root .tmp/pycodex_storage --host 127.0.0.1 --port 5577
Upload the default Codex home, or specify another directory:
pycodex --put @127.0.0.1:5577
pycodex --put /path/to/.codex/@127.0.0.1:5577
After verifying a download round trip, the command prints a reusable
pycodex --call ... command. Run it on the target machine using a storage
address reachable from that machine.
Development and reference
uv sync --dev
uv run pycodex
uv run --dev pytest
- Test environments and coverage
- Repository development rules and reading guide
- Architecture, package layout and runtime contracts
- Prompts, configuration and context
- Upstream alignment
- Responses compatibility server
- Portable configuration contracts
See the 0.6.0 release notes for the major changes and
Python integration migration from 0.5.0. Implementation details live in docs/.
Markdown view modes
Markdown file views include a Show source / Show preview button. Switching modes preserves each mode's reading position without replaying a previous line target.
Responses reliability
Providers can opt into ordered failover with fallback_models = ["backup-one", "backup-two"] in their [model_providers.<name>] table. Each target uses the
same endpoint, credentials, history and tools; request settings are rebuilt for
that model, while the session's selected model stays unchanged. Transient errors
exhaust stream_max_retries before advancing. Unavailable-model errors and
first-data timeouts advance immediately. Existing retry notices name both models
and include the underlying error.
Failed models cool down for 900 seconds within that client. Subsequent requests
skip them, retry them after expiry, and restart the full chain when every target
is cooling. fallback_cooldown_seconds = 0 disables cooldown; explicit model
selection clears it. New sessions and subagents have independent cooldowns.
An absent or empty list disables failover. Invalid input, context-length and
terminal incomplete errors retain their existing handling. Backups must support
the active tools and context size; compaction keeps the selected model's policy.
Responses HTTP 408/429 and server errors use the provider retry budget.
Rate-limit messages containing try again in or retry after supply the retry
delay. CLI and workspace retry notices include the underlying error details;
workspace notices survive browser refresh and active-turn reattachment, and
remain separate from the assistant reply. They are not persisted in rollouts.
Workspace input history
The workspace message box supports Up/Down input history. Up recalls earlier human inputs in the current conversation; Down returns to the unsent draft. Automatic wake-up messages are excluded, and arrows retain normal cursor movement inside multiline drafts.
Metadata
Release files for python-codex 0.6.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| python_codex-0.6.0.tar.gz | 328.4 kB | With provenance | Details for python_codex-0.6.0.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| python_codex-0.6.0-py3-none-any.whl | Python 3 | none | any | Details for python_codex-0.6.0-py3-none-any.whl |
Total release size: 689.4 kB
Release files / python_codex-0.6.0.tar.gz
| Download URL | python_codex-0.6.0.tar.gz |
|---|---|
| Size | 328.4 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
f97e39f3f11a3544534a5e112769dcd56236a0cc3fec53e05ca127dc5fc19fd0
|
|
BLAKE2b-256 checksum How to use checksums |
04e6321f83cfb656bb60bf709e8483e92729fa08c70b0eacf64bc8a29c36c44f
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Oct 31, 2026.
Transparency logRelease files / python_codex-0.6.0-py3-none-any.whl
| Download URL | python_codex-0.6.0-py3-none-any.whl |
|---|---|
| Size | 361.0 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
508bb13121de45f8bfb0908181365071773e883c8a145ac42ea4f4b3f5cb6658
|
|
BLAKE2b-256 checksum How to use checksums |
5865e279c0183b5bb6275bf49943940fc6bbb85912cf00b6ac893a64e1e25f98
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Oct 31, 2026.
Transparency logSecurity for python-codex 0.6.0
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
Full provenance
Both files in this release have been uploaded with provenance metadata.
Attestation for python_codex-0.6.0.tar.gz
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Oct 31, 2026.
Transparency logA note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.