Skip to content

Latest commit

 

History

History
347 lines (240 loc) · 22.4 KB

plugins.md

File metadata and controls

347 lines (240 loc) · 22.4 KB

Optional agent tools and plugins

Run these commands inside Agentainer as the regular container user. Choose the plugins you want for each agent. The image installs none of them, at build or at startup. Its built-in agent settings are described in default configuration, and where each agent keeps settings and skills is in setting up the agents.

Before you install anything from this page:

  • None of these tools is tested against the currently pinned agents. Verification limits lists what was checked and when.
  • Try a plugin in a separate test home first. Plugins and hooks change what an agent runs or sees.
  • After installing, inspect your settings, hooks and instruction files for changes, restart the agent, and check that the plugin's tools or hooks actually run. An install success message does not prove the agent loaded it.
  • Keep the same persistent home to keep home-based plugin settings and files. A persistent home does not keep software installed elsewhere in the container, and external dependencies and account connections need their own setup.
  • Do not copy upstream example instructions over your project's or the shared AGENTS.md.

Contents

Recommended starting points

For Codex, OpenCode and Antigravity, Addy Osmani's agent-skills provides these reusable workflows:

Skill Use
code-simplification Reduce complexity while preserving behavior
code-review-and-quality Review correctness, readability, architecture, security and performance
frontend-ui-engineering Build responsive, accessible interfaces
security-and-hardening Review trust boundaries, authentication, untrusted input and secrets

Start with simplification and review. Add frontend and security workflows for the work you do. Google uses the review skill in its Antigravity code-review tutorial. These are community-maintained skills, not agent-vendor defaults.

Superpowers is an optional workflow package for all four agents. It adds planning, test-driven development, debugging and verification skills. It also directs agents to invoke skills at the start of work, so expect broader workflow changes than a single cleanup skill. Choose it separately from the focused skills above. Follow your project's instructions and approval requirements when a workflow proposes commits or other actions.

For more visual design guidance, UI UX Pro Max provides palettes, typography and design-system tools. Its installation guide supports Codex, OpenCode and Antigravity. Keep its scripts and data with the skill; copying only its instructions is insufficient.

Token and context tools

These tools solve different problems. Start with one, measure the result, and keep the original command output available when checking a failure.

Tool Purpose Suggested use
RTK Filter command output before the model reads it First trial for output-heavy builds and tests; enable each agent integration separately
context-mode Execute analysis through MCP tools and search indexed content Optional trial for large logs, research and long sessions
Superpowers Planning, test-driven development, debugging and verification workflows Choose for its development process; it is not a token compressor
Context7 Retrieve library documentation and examples Optional documentation lookup through its CLI or hosted MCP server
Neuledge Context Query a locally curated documentation collection Optional local alternative to Context7

Do not enable every tool across every agent as one setup step. RTK and context-mode can both intercept tools, while Superpowers changes the workflow. Check them independently before combining them. No token or billing savings from these tools have been measured in Agentainer.

RTK

RTK filters shell-command output. Its upstream guide distinguishes reduced shell output from total input tokens or billing savings; its token counts are estimates. The current editions include an Apache-2.0 rtk-oss build and a full build with additional licensed components. Start with the OSS edition for command filtering. Both editions install a binary named rtk; the API proxy is a separate component.

Use upstream release binaries, not the unrelated package with the same name on crates.io. The installer source supports RTK_EDITION=oss and an exact RTK_VERSION, checks the published SHA-256, and installs under ~/.local/bin by default. Replace vX.Y.Z with an actual reviewed release tag before running these commands inside the container:

installer=$(mktemp)
curl -fsSL https://raw.githubusercontent.com/rtk-ai/rtk/refs/heads/master/install.sh -o "$installer"
less "$installer"
RTK_EDITION=oss RTK_VERSION=vX.Y.Z RTK_INSTALL_DIR="$HOME/.local/bin" sh "$installer"
rm -f "$installer"
export PATH="$PATH:$HOME/.local/bin"
rtk --version
rtk init --help

Review the downloaded installer before executing it. Installing the binary does not configure every agent. The upstream agent table documents these separate setup commands; choose only the integration you want:

Agent Upstream setup command
Claude Code rtk init -g
Codex rtk init -g --codex
OpenCode rtk init -g --opencode
Antigravity CLI rtk init -g --agent antigravity

RTK's agent table describes Codex command rewriting through updatedInput, while context-mode's Codex guide describes deny-only PreToolUse routing. Those upstream descriptions disagree. Actual rewriting behavior with the pinned Codex is not verified.

An upstream minimum version for older OpenCode does not show compatibility with OpenCode v2. rtk gain reports RTK's savings estimates. See the installation guide for removal.

context-mode

context-mode provides MCP execution and indexing tools, plus client-specific routing hooks. It stores session and indexed content locally. Its current license is Elastic License 2.0. Review its configuration and retention behavior before indexing sensitive files.

For an MCP-only trial, select a reviewed exact package version from published npm metadata. Replace VERSION below; do not run the placeholder literally:

npm install --global 'context-mode@VERSION'
context-mode doctor

Merge this server into the existing mcpServers object in ~/.agentainer/mcp.json, preserving its other entries:

{
  "mcpServers": {
    "context-mode": {"command": "context-mode"}
  }
}

Stop agents that write their config files, then run agentainer-mcp sync as described in MCP synchronization. MCP-only registration exposes tools; it does not install routing hooks. Use the upstream installation guide for each client's full plugin setup, restart steps, diagnostics and removal. Codex's upstream setup includes hook feature flags and hook trust, and the Antigravity CLI setup differs from the Antigravity desktop IDE.

Context7

Context7 retrieves library documentation and code examples. It is a documentation service, not a general command-output compressor. Its hosted MCP server avoids installing a local MCP server package. Queries go to the hosted service, so use public library names, versions and focused documentation questions rather than secrets.

For the upstream setup flow, select an exact reviewed ctx7 version from published npm metadata and replace VERSION:

npx 'ctx7@VERSION' setup

The setup guide documents the Node.js requirement, OAuth flow and choice between CLI skills and MCP. To use Agentainer's shared MCP configuration instead, merge the following entry into its existing mcpServers object and provide CONTEXT7_API_KEY through the runtime env file:

{
  "mcpServers": {
    "context7": {
      "url": "https://mcp.context7.com/mcp",
      "headers": {"Authorization": "Bearer ${CONTEXT7_API_KEY}"}
    }
  }
}

This endpoint and bearer-header format come from the upstream client guide. Then run agentainer-mcp sync (MCP synchronization). The sync does not give Antigravity headers that contain ${NAME}. For Antigravity, add the server by hand under a name that mcp.json does not use, because the next sync rewrites a server it manages (agents).

After restarting the client, check that resolve-library-id and query-docs are available and complete a public-library lookup. Request documentation for the project's actual version and check the original source when accuracy matters. Remove generated setup with the matching ctx7 CLI's remove command, or remove the server entry and sync again. See the upstream CLI guide for the selected version's options.

Neuledge Context: local option

Neuledge Context provides a local documentation collection you can tailor to a project. Its npm package is @neuledge/context, and its executable is context.

Choose an exact reviewed version from published npm metadata, replace VERSION, and run inside the container:

npm install --global '@neuledge/context@VERSION'
context add /workspace --path docs --name project-docs
context list
context query project-docs 'configuration'

The upstream CLI reference documents these commands. Adjust the source path and query to your project. It also supports Git repositories, websites and prebuilt registry packages. Packages reside in ~/.context/packages/ on the persistent home volume. Re-run context add when source documentation changes.

For a collection limited to those installed project docs, merge this entry into the existing mcpServers object in ~/.agentainer/mcp.json:

{
  "mcpServers": {
    "neuledge-context": {
      "command": "context",
      "args": ["serve", "--libs", "project-docs"]
    }
  }
}

The upstream serve reference documents stdio transport and --libs, which restricts available packages and hides registry search/download tools. Installed documentation queries run locally; installation, registry discovery and remote-source imports require network access. The agent still sends retrieved content to its configured model.

Then run agentainer-mcp sync (MCP synchronization), restart the agent and check a get_docs lookup.

Claude Code

These plugins are available through Anthropic's official marketplace. Superpowers is maintained by its own project:

Plugin Purpose
code-simplifier Simplify code while preserving its behavior
frontend-design Help implement web interfaces
security-guidance Add security checks through hooks and reviews; can make additional model calls
Superpowers Planning, test-driven development, debugging and verification workflows

Check registered marketplaces:

claude plugin marketplace list

If claude-plugins-official is missing, add it:

claude plugin marketplace add anthropics/claude-plugins-official

Install each plugin you choose with its own command. User scope applies to your home across projects. The Superpowers command comes from the project's Claude installation guide:

claude plugin install code-simplifier@claude-plugins-official --scope user    # simplification
claude plugin install frontend-design@claude-plugins-official --scope user    # web interfaces
claude plugin install superpowers@claude-plugins-official --scope user        # broader development workflow
claude plugin install security-guidance@claude-plugins-official --scope user  # security hooks and reviews

List installed plugins, then start a new Claude session to load them:

claude plugin list

For example, disable or uninstall code-simplifier with:

claude plugin disable code-simplifier@claude-plugins-official --scope user
claude plugin uninstall code-simplifier@claude-plugins-official --scope user

See Claude's installation and scope documentation. Avoid project scope when the choice should apply only to you.

In the 2026-10-21 check, the Claude installer rewrote the seeded permission mode from manual to default and kept the other seeded settings. Inspect ~/.claude/settings.json after installing a plugin.

Codex

Browse configured marketplaces and their available plugins:

codex plugin marketplace list
codex plugin list --available --json

Select a plugin from that output. Replace PLUGIN and MARKETPLACE below with its actual names; they are placeholders:

codex plugin add 'PLUGIN@MARKETPLACE'
codex plugin list

Remove that installation with:

codex plugin remove 'PLUGIN@MARKETPLACE'

To install the full Addy Osmani skills pack, follow its Codex installation guide:

codex plugin marketplace add addyosmani/agent-skills
codex plugin add agent-skills@agent-skills

This installs the pack, including workflows beyond the four starting skills. The project's Codex setup guide explains skill invocation and discovery. Do not paste its meta-skill into your shared instructions; Codex discovers the installed skills on demand.

For Superpowers, open /plugins in Codex, search for Superpowers, and select its installation option (project guide). For dedicated security scans, the official Codex Security guide documents the same browser flow for Codex Security. Start a new session after installation.

Agentainer does not enable these recommendations automatically. See the official Codex CLI plugin reference for supported sources and marketplace management.

OpenCode

Not verified: whether OpenCode loads installed plugins and skills. In the 2026-10-21 check, plugin add wrote the expected plugins entry and plugin remove removed it, but plugin list and the native plugin API returned empty lists, and the skill API returned an empty list after the four recommended skills were copied in.

List installed package plugins:

opencode plugin list

Choose a package after reviewing its publisher, source and required configuration. Replace PACKAGE@VERSION with the package's actual name and an exact published version:

opencode plugin add 'PACKAGE@VERSION'
opencode plugin list

The installer adds a global package plugin. Remove it using the same configured package specifier:

opencode plugin remove 'PACKAGE@VERSION'

OpenCode also supports local plugin files and directories. Its configuration key is plugins, plural. See the OpenCode plugin documentation for package sources, configuration and reload behavior.

Keep Agentainer's existing ~/.config/opencode/opencode.json when editing configuration. Adding opencode.jsonc prevents Agentainer's MCP sync from updating OpenCode; see MCP synchronization.

For the focused starting skills, follow the agent-skills OpenCode setup guide. Select the desired skill directories and retain their shared references. The guide notes that individual skill installs can omit required pack-level references. Do not replace your project's AGENTS.md with the upstream repository's instructions.

Superpowers documents an OpenCode V2 installation. Merge this entry into the existing plugins array in your home configuration, preserving other settings:

{
  "plugins": ["superpowers@git+https://github.com/obra/superpowers.git"]
}

The JSON above shows the setting to merge, not a replacement configuration file. Restart OpenCode after configuring it. This Git source follows upstream changes; select a reviewed immutable revision for a fixed installation, using the source forms in the OpenCode plugin documentation.

Antigravity

In an interactive agy session, use /plugin to browse the official marketplace. From a container shell, list installed plugins:

agy plugin list

Replace PLUGIN with the name you selected:

agy plugin install 'PLUGIN@antigravity-plugins-official'
agy plugin list

Disable or uninstall it with:

agy plugin disable PLUGIN
agy plugin uninstall PLUGIN

The agent-skills Antigravity guide documents installation of its full pack:

agy plugin install https://github.com/addyosmani/agent-skills.git

That guide records limitations with legacy short command wrappers in some Antigravity releases. Use the installed namespaced skills directly and check discovery in your session. Whether our pinned CLI has that wrapper limitation is not verified.

For Superpowers, its Antigravity installation guide documents:

agy plugin install https://github.com/obra/superpowers

Choose each pack independently. See Antigravity's plugin documentation for discovery, installation and account requirements. Compatibility of other Claude plugins with Antigravity is not verified.

Verification limits

These checks ran on 2026-10-21 on arm64 with disposable homes and the agent versions pinned at that time. No credentials were mounted and no model requests were sent. No public result artifact records them, and they have not been repeated with the current pins in versions.env, so treat them as not verified.

Agent or package Observed result
Claude Code All four recommended plugins installed from the official marketplace, appeared enabled in claude plugin list --json, and disappeared after uninstalling them. The installer normalized the permission-mode spelling as described above.
Codex Superpowers and agent-skills installed from local checkouts of their upstream repositories, appeared enabled in codex plugin list --json, and disappeared after removal. The unauthenticated default catalog returned no available entries, so the curated Superpowers and Codex Security browser installations remain not verified.
Antigravity Both packs installed from local upstream checkouts. agy plugin list --json listed the four recommended agent-skills workflows and Superpowers' skills; both packs disappeared after uninstalling them.
OpenCode Superpowers' pinned Git package entry was added and removed successfully. Runtime plugin and skill discovery remain not verified, as described above.
UI UX Pro Max The ui-ux-pro-max-cli installer created per-project assets for Codex, OpenCode and Antigravity. Its local ui-ux-pro-max/scripts/search.py design-system generator succeeded for all three layouts. Agent-driven use remains not verified.

The pack checks used Superpowers revision b739a8a and agent-skills revision aeac41d. The UI UX Pro Max check used the npm CLI release 2.15.0. That installer also included auxiliary design skills; inspect its generated files before adding them to a project.

Verified: the claude plugin, codex plugin, opencode plugin and agy plugin subcommands used on this page exist in the pinned CLIs (--help in the current image). Package names, sources and setup steps come from the linked upstream guides. Not verified: coding workflows, account connections, model costs, hook execution, approval prompts and plugin effectiveness. None of these checks is a security audit.

The RTK, context-mode, Context7 and Neuledge Context examples were checked against upstream documentation on 2026-10-22. Not verified in Agentainer: their installers, authentication, agent hooks, offline behavior and savings.