Code with Codex
Codex Responses includes stable prompt-cache keys and cache-preserving Astra effort changes.
Use the codex pack to start fast-agent with a coding agent, a
Codex-optimised filesystem search sub-agent, Responses-family transport, and
model-selected filesystem editing tools.
This starts fast-agent pre-configured for a Codex-style coding workflow.
What the codex pack gives you
- A
devcoding agent for interactive software work - A bounded rg-first search helper backed by
codexspark - WebSocket-capable transport for modern Codex/OpenAI models
- Filesystem editing tools selected for the model: Astra defaults to
write_text_fileplusedit_file; patch-oriented models useapply_patch - Preconfigured MCP targets available from
/mcp attach
The coding agent has a minimal system prompt plus tools for the shell,
filesystem and fast-agent services. AGENTS.md is included automatically if
present. Customise the agent by editing .fast-agent/agent-cards/dev.md.
Astra's writer/editor default applies to both Responses and Codex Responses,
including the astra and codexplan aliases. To explicitly select the
Codex-style patch interface instead:
When apply_patch tool calls are previewed in the console, large patches are
collapsed with a (+N more lines) tail. You can tune or disable that limit via
logger.apply_patch_preview_max_lines in fast-agent.yaml, or interactively
with fast-agent config display.
Use /skills to discover, add, remove and update skills. Use /connect to
connect to MCP Servers.
Codex authentication
If you want to use the Codex OAuth models directly, authenticate once first:
Login uses device auth by default: open the displayed URL in a browser and enter the one-time code. No local callback server is required. Only enter a code from a login you started yourself.
If device auth is unavailable for your account, use the existing browser callback flow instead:
Device login failures do not automatically switch to browser login. Credentials are stored in your OS keyring, with a secure file fallback. After that you can use Codex OAuth model aliases such as:
codexplan— GPT-6-Astra with medium reasoning
If you prefer, you can also run model setup explicitly:
Web search
Use fast-agent go --model 'astra?web_search=true' to enable automatic web_run
on Codex Lite without shell access; use web_search=false to disable it.
In a running conversation, /model web_search on and /model web_search off
control the same feature. Sol and public Responses keep their hosted search route.
See standalone web search
for configuration, the internal endpoint caveat, and a runnable library example
with caller-supplied authentication.
Installation
fast-agent requires Python 3.12 or above. Install with:
Or a specific version of Python:
This installs the fast-agent executable.
Next steps
From the fast-agent prompt:
- Use
/skillsto view and manage skills - Use
/mcp attach <name>to attach preconfigured MCP servers; use/connectfor ad-hoc targets - Ask the agent to create additional cards in
.fast-agent/agent-cards/ - Switch agents with
@ - Configure compaction, hooks or automation with the available skills