agent-toolbox

agent-toolbox provides three portable plugins for AI coding agents:

• SWE — skills for sharpening and planning work, implementing and reviewing changes, and shipping pull requests.
• Lab — source-backed research, reproducible experiment loops, and data-visualization guidance.
• Uni — interactive lessons with questions, active recall, and exercises that deepen understanding. See Uni installation and usage.

It also provides Claude Code-only mods under claude-mods/:

• activity — a live pane of the session's tool calls, with durations and failures. Install it with claude plugin install activity@agent-toolbox, then run /activity to open it.
• subagents — an animated pane of the session's subagents: a Clawd sprite per agent with its type, task, latest tool call, and elapsed time. It opens when the first subagent starts, where the screen has room for a side pane. Install it with claude plugin install subagents@agent-toolbox; /subagents shows or hides it.

SWE treats the merged pull request and the code as the record. Large work can start from a short plan that hands off to a fresh session, with configured mirroring for Linear-linked plans. Its implementation discipline challenges unnecessary code and simplifies the completed change before final verification. Sound work can stay unchanged. Publication follows the user's authorization.

Lab keeps source checks and experiment boundaries while scaling research artifacts and delegation to the question. Neither plugin requires a scheduler or a separate workflow service.

Uni keeps the learner's attempts, explanations, and review outcomes central. It combines guided reconstruction, criticism, examples, retrieval, and transfer as needed. Reviews run when invoked. Anki or another external tool can handle long-term scheduling.

Install

Claude Code

/plugin marketplace add kpeez/agent-toolbox
/plugin install swe@agent-toolbox
/plugin install lab@agent-toolbox
/plugin install uni@agent-toolbox

Codex CLI

codex plugin marketplace add kpeez/agent-toolbox
codex plugin add swe@agent-toolbox
codex plugin add lab@agent-toolbox
codex plugin add uni@agent-toolbox

Optional skills.sh skill-only install

npx skills@latest add kpeez/agent-toolbox

This installs editable skills only. Plugin agents still require a plugin installation.

Teaching workspace

Uni includes teach for a persistent learning workspace with HTML lessons, references, a mission, and learning records. It preserves the original workflow and four format templates, with upstream attribution. Use learn for conversational tutoring and explicitly invoke teach to author a learning workspace: $uni:teach in Codex or /uni:teach in Claude Code.

An earlier installation may have a ~/.codex/skills/teach or ~/.agents/skills/teach symlink to this repository's former skills/teach directory. After updating Uni, remove that legacy symlink to avoid duplicate discovery. Preserve any independently edited skill directory.

Delegation

The implementation skill routes substantial work with independent parts to swe:orchestrate. To make the same decision before skill selection, add this rule to your user-level AGENTS.md (Codex) or CLAUDE.md (Claude):

For substantial work with independently executable parts, use swe:orchestrate to delegate them. Keep tightly coupled work with one agent. Send a single independent review directly to the reviewer.

The plugin does not edit user instructions. Existing sessions may retain prior instructions; use a fresh session after updating.

Project documents

ADRs, notes, and research default to .agents/docs/. Follow an explicit project or user override. Create subdirectories as needed; the location may be a regular directory or symlink, tracked or ignored. No setup step, vault, or generated AGENTS.md/CLAUDE.md is required.

Plans

Plans live in <main checkout>/.agents/plans/, shared by all worktrees. The directory keeps itself out of git with its own .gitignore.

A plan named ABC-123-short-slug.md can be mirrored one way to a document on Linear issue ABC-123 by a configured host Stop hook (plan_sync.py). It reads a Linear personal API key from LINEAR_API_KEY, or else from ~/.config/swe/linear-api-key (mode 0600, honoring XDG_CONFIG_HOME), so no shell configuration is needed. Without a key, or without an issue key in the name, plans stay local and nothing is blocked. Hook delivery varies by host; confirm sync state when it matters. GitHub-linked plans remain local unless a separate publication route is configured.

New autoresearch runs keep their program, ledgers, logs, configuration, and generated figures under artifacts/autoresearch/<tag>/. Existing runs retain their recorded paths. A configured vault receives closeout notes and only the reference figures those notes embed.

Layout

• plugins/swe/ — SWE skills, agents, and the plan-sync hook.
• plugins/lab/ — Lab skills and their runtime scripts and references.
• plugins/uni/ — tutoring skills, the teaching workspace workflow, templates, concept exercises, and research notes.
• claude-mods/<mod>/ — Claude Code function-hook mods, one plugin per mod, listed in the Claude catalog only.
• .claude-plugin/marketplace.json — Claude marketplace catalog.
• .agents/plugins/marketplace.json — Codex marketplace catalog.

Versioning

Update each plugin's version manually in both manifests, and keep the values identical (a mod under claude-mods/ has the Claude manifest only):

• plugins/<plugin>/.claude-plugin/plugin.json
• plugins/<plugin>/.codex-plugin/plugin.json