Think before AI thinks.
Axiom is a Codex plugin with a collection of focused skills for turning requests into clear, actionable work. It helps clarify intent, plan tasks, maintain repository instructions, research, review decisions, and handle actions that need explicit authorization or a recovery plan.
Describe what you want in Codex, or name a skill directly. Axiom loads the smallest matching skill set and its relevant references; everyday requests continue through Codex normally. The skills use Codex's existing tools and permissions, with no background service or automatic updater.
Supported host: Codex. Claude Code support has been retired; see
Deprecated Skills and Support.
Install | Update | Use a skill | Documentation
Run these commands in a terminal to add the marketplace and install Axiom:
codex plugin marketplace add wheakerd/axiom --sparse .agents/plugins
codex plugin add axiom@axiomThe marketplace selects an immutable published version. Development commits
on main do not advance that selection.
Start a new Codex chat or CLI session. Open /hooks and compare the installed
entry with the Hook Reference before trusting it.
Stop if the definitions differ. The Getting Started guide
owns the full installation and first-use procedure.
Refresh the installed marketplace snapshot from a terminal:
codex plugin marketplace upgrade axiomIn a supported Codex workspace plugin UI, use Refresh. After updating,
start a new session and inspect /hooks again. Automatic update behavior is
controlled by Codex and your settings. See Managing an Installation
for disabling, removal, and troubleshooting.
For a marketplace clone timeout, follow the guide's
recovery procedure.
In Codex, try a read-only request that matches a skill:
Use $agents-architect to audit this repository's AGENTS.md instruction system.
Report findings only; do not modify files.
Then try an ordinary request:
Summarize the purpose of this README. Do not modify files.
The first request should select agents-architect; the second should continue
normally without another skill. When the startup gate is enabled or you
explicitly invoke using-axiom, its
optional notice guidance
leaves explanations and layout to the model and host under your output-verbosity
preferences. Explanations identify Axiom skills as Axiom:<skill-name> and
external skills (including repository-local skills), plugins, or MCP capabilities
by their own inline-code names,
with a brief reason. A quiet response is valid. These are expected behaviors,
not a claim that your installation has already reproduced them.
Important
A request that matches no Axiom skill still follows Codex permissions and repository instructions. No match is not a safety certification. Selecting a skill does not grant permission to edit, publish, deploy, or delete anything.
Use natural language or include $skill-name in a Codex message. The examples
below are chat prompts, not terminal commands. Each linked source describes
the skill's full scope and requirements; Examples gives
more requests and expected behavior.
Link: Startup routing skill
Description: Finds the smallest matching skill set and leaves unmatched work with Codex. It loads through the startup hook.
Usage:
Use $using-axiom to identify the skill that fits this request.
Link: Request clarification
Description: Resolves materially different interpretations with a focused question and useful options.
Usage:
Use $clarify-intent to clarify what "clean up this project" should mean
before making changes.
Link: Bounded task delegation
Description: Assesses simple work against your ordered model preferences
while preserving the main model. Requires the active using-axiom gate,
verified Ultra reasoning, and your consent to invoke the skill.
Usage:
I authorize $delegate-simple-task to assess this typo fix using my model
preferences, only with the using-axiom gate active and my Ultra reasoning verified.
Link: Task planning and revision
Description: Turns requirements into steps, dependencies, and acceptance criteria, or updates an existing plan.
Usage:
Use $task-planning to plan this settings-page change.
Do not implement it yet.
Link: Supervised progress across sessions
Description: Explicitly invoked progress through fresh workers and coordinator handoffs, with a verified native supervisor. Stops after verified completion or cancellation; local scheduling requires Codex and the machine to remain available.
Usage:
Use $continuous-progress to finish this task. Create fresh execution and
successor chats as needed, register an independent native supervisor,
and stop its schedule when the task is verified complete or I cancel.
Link: Research through a local browser or client
Description: Limits information disclosed in requests and handles site challenges through a user handoff.
Usage:
Use $local-web-search to find official migration guidance
through my local browser.
Link: Repository instruction architecture
Description: Audits and maintains AGENTS.md, routed .agents/ guidance,
and repository skills.
Usage:
Use $agents-architect to audit this repository's instruction system.
Report findings only.
Link: Packaged plugin architecture
Description: Designs or audits skills, routes, manifests, hooks, and compatibility evidence for a packaged agent plugin.
Usage:
Use $agent-plugin-architect to audit this Codex plugin's package architecture
without editing it.
Link: Codex usage analysis
Description: Finds ways to reduce context, token, and tool overhead while preserving required quality.
Usage:
Use $optimize-codex-usage to review this workflow's context overhead
and recommend reductions.
Link: Axiom task review
Description: Reviews observable routing, authorization, actions, and evidence for an identified task.
Usage:
Use $review-axiom-task to review why Axiom paused the previous task.
Do not retry it.
Link: Consequential external actions
Description: Checks the actor, destination, exact content, and existing authorization before execution and verification.
Usage:
Use $confirm-external-action to prepare this announcement for publication.
Show the exact destination and text; do not publish yet.
Link: Traceable Git work
Description: Handles explicit checkpoints, consolidation, recovery, and submissions that need traceability.
Usage:
Use $traceable-git-submit to create a checkpoint of the README changes.
Do not push.
Link: Changes with recovery planning
Description: Plans, rehearses, or executes persistent changes with verified recovery and outcome checks.
Usage:
Use $reversible-system-change to plan this database migration and its rollback.
Do not execute it.
No current top-level bundled skill is marked deprecated. Earlier nested
AGENTS Architect topics are now supporting references within
agents-architect, not separate skills to
install or invoke.
Claude Code installation and runtime support was removed in
v0.11.0. Existing users should disable or uninstall
their old Axiom installation in that host and use the Codex installation above.
Updating the repository does not remove cached installations.
agent-plugin-architect can still help design Claude Code plugins for other
projects. That skill does not make Claude Code a supported host for Axiom.
The release badge follows GitHub Releases automatically. The other status badges link to the repository's live workflow results:
- Distribution and publication guards validate the package, documentation, and repository publication rules.
- CodeQL scans Python and GitHub Actions workflows for potential vulnerabilities. Its configuration is in the security workflow.
- MIT identifies the project's license.
A workflow badge reports a run result, not a guarantee that the project is free of vulnerabilities. Axiom is a public beta, and its instructions do not replace Codex permissions or provide a sandbox. Read the Trust Model for the security boundaries and SECURITY.md to report a vulnerability privately.
- Documentation index: guides and references by task.
- Compatibility: supported hosts and known limitations, backed by the release-status record.
- Architecture: how routing and skills fit together.
- Changelog: changes and migration notices.
- GitHub Issues: help and bug reports.
- Contributing: contribution and validation guidance.
MIT.