@axsept/codex

One-line SkillChain integration for OpenAI Codex CLI. Captures the skills you demonstrate while working with Codex and builds your verified SkillChain profile. No conversation content leaves your machine — only anonymized skill signals.

It runs on the same open standard as the Claude Code plugin: the same local @axsept/mcp-server, the same @axsept/event-schema, and the same local @axsept/anonymizer. Codex is metadata (source.provenance: "codex"), not a new pipeline.

Install

npx @axsept/codex init

That's the whole thing. init:

1. Opens your browser to skillchain.tech/connect?source=codex — sign in, click Authorize Codex CLI, copy the sk_user_… key.
2. Prompts you to paste the key, and writes it into ~/.codex/config.toml under a clearly-marked SkillChain block (file mode 0600).
3. Registers the SkillChain MCP server and adds the activity-tracking instructions to ~/.codex/AGENTS.md.

Restart Codex (or start a new session) and your work flows to your profile. Existing config.toml / AGENTS.md content is preserved — only the SkillChain-managed block is edited, and re-running init replaces it in place (never duplicates).

Commands

| Command | What it does | |---|---| | npx @axsept/codex init [--with-hooks] [--base-url URL] [--key KEY] | Set up MCP + instructions (+ optional auto-capture hooks) | | npx @axsept/codex connect [--base-url URL] | Re-authorize / rotate your API key | | npx @axsept/codex status | Show what's wired up (MCP, instructions, hooks, base URL, key) | | npx @axsept/codex uninstall | Remove every SkillChain block + hook script |

What this does

Once installed, the bundled activity-tracking guidance (in your AGENTS.md) tells Codex to call the SkillChain MCP tools at meaningful moments:

• log_activity — records a discrete activity you demonstrated a skill on. Called sparingly (2–8×/hour of focused work), not per message.
• record_session_summary — at natural boundaries, captures your directing / orchestration skill across the whole session.
• get_profile_snapshot — read-only; lets Codex see your current signature to avoid re-logging and personalize.

Every activity is tagged source: codex so you always know which tool produced each signal.

Automatic capture (optional, experimental)

npx @axsept/codex init --with-hooks

This additionally installs Codex Stop / PreCompact lifecycle hooks that nudge the model to log after substantial work (the same idea as the Claude plugin's hooks). They ship in dry-run by default and are still being calibrated against Codex's hook contract — notably the transcript_path format and the decision:block re-prompt semantics, which we verify against a live Codex install before flipping them on. Until then they're safe no-ops. The rich path above (MCP tools + AGENTS.md) is the supported way to capture today.

Hook controls (env vars, all optional): SKILLCHAIN_HOOKS_DISABLED=1 (kill switch), SKILLCHAIN_LOG_DRY_RUN=0|1, SKILLCHAIN_LOG_VERBOSE=1, and the SKILLCHAIN_STOP_* thresholds documented in the Claude plugin.

What this does NOT do

• It does not send your conversation content to SkillChain. The local @axsept/anonymizer strips PII and secret patterns (emails, phone numbers, API keys, SSH keys, JWTs, …) from every text field before any network call.
• It does not run a separate LLM or charge you for inference. Skill classification is done by the Codex model as part of the conversation you're already having.
• It does not retain your conversation locally beyond what Codex already does.

Privacy and trust

• All text fields are anonymized locally before transmission. The pattern set is open source.
• The integration only talks to https://skillchain.tech (or whatever --base-url you set).
• The MCP server source is open (@axsept/mcp-server). npm provenance attestations (verifiable with npm view @axsept/mcp-server --json | jq .provenance) are added once the integration packages move to their public source repo — they are not generated during the private-repo beta.

How it's built (for contributors)

This package is a thin, Codex-specific shell. The reusable pieces live elsewhere:

• Capture/emit engine: @axsept/mcp-server (unchanged, run via npx).
• Shared instruction layer + hook libs: @axsept/wrapper-core — the host-neutral SKILL.md / AGENTS.md and hooks/lib/{state,config}.sh are vendored from there and guarded against drift by a sync test.
• Codex-specific glue (the only original code here): lib/config-toml.mjs + lib/agents-md.mjs (idempotent marker-block editors), lib/connect.mjs (the browser auth flow), bin/cli.mjs, and the hooks/ adapters (decision:block emit contract, Codex transcript.sh).

Source: packages/skillchain-codex/. Apache-2.0.