Install and manage Agent Plugins 1.0 across your AI agents with one CLI.
Table of contents
- Use plugins
- What the CLI does
- Restore your plugin setup
- Any Agent Plugins 1.0 package
- Supported clients
- Already built with UAP SDK
- Find and verify plugins
- Lifecycle safety
- More commands
- Build plugins
- Contributing
- License
Keep a reproducible set of plugins and dependencies, and carry it to another computer. The manifest records what you need; the lockfile pins the exact packages to restore. This workflow ships in the published CLI; see Restore your plugin setup. For the published installer, start with the Quick start.
Our installer automation principle is to complete every ownership-safe client step automatically and ask the user only for genuine external approvals, authentication, or reloads.
Read the documentation · Browse plugins
Use plugins
Install, inspect, update, repair, and remove packages with the existing installer. The Use guide source collects this journey. See Build plugins below to create and check a package.
Quick start
With Node.js 22+, run:
npx universal-agent-plugins add context7
The CLI automatically finds compatible agents, including Codex and Claude Code. Choose agents if several are found; a single detected agent is selected automatically. Follow the printed setup steps, then start a new agent session.
With Homebrew:
brew install 777genius/agentplugins/agentplugins
agentplugins add context7
Without Homebrew, use the native installer:
curl -fsSL https://raw.githubusercontent.com/777genius/universal-agent-plugins/main/install.sh | sh -s -- add context7
Install the matching native binary and add your first plugin:
curl -fsSL https://raw.githubusercontent.com/777genius/universal-agent-plugins/main/install.sh | sh -s -- add context7
Homebrew also works on Linux: run brew install 777genius/agentplugins/agentplugins,
then agentplugins add context7.
Run in PowerShell:
irm https://raw.githubusercontent.com/777genius/universal-agent-plugins/main/install.ps1 | iex
& "$HOME\.local\bin\agentplugins.exe" add context7
To install the latest installer globally with Node.js 22+:
npm install -g universal-agent-plugins
Homebrew and the installers select the native binary for your OS and architecture.
The installer scripts verify its published SHA-256 and reported version, then
replace the CLI atomically. They install into $HOME/.local/bin unless
AGENTPLUGINS_BIN_DIR is set.
The native CLI does not require Node.js. The npx option requires Node.js 22+
and downloads and runs the verified Go binary. Individual plugins may have
their own runtime requirements, which the CLI checks before installation.
The plugin package is downloaded and verified once, then prepared for each selected agent.
What the CLI does
- installs one plugin in one or several supported agents;
- updates, repairs, and removes only files it manages;
- converts the same Agent Plugins 1.0 package into each agent's native format;
- keeps activation and OAuth prompts visible to you.
Restore your plugin setup
These commands are part of the published CLI (agentplugins 0.1.90 and later).
The Quick start above covers installation.
On the source computer, open a folder to use for the transfer and run:
agentplugins deps export
Copy that folder to the new computer, install the required target clients, and open the copied folder. Then restore the recorded plugins and dependencies:
agentplugins deps import --install
The optional directory argument defaults to the current folder. By default,
export reads the global user environment and import writes it. Only uap.toml and uap.lock
are transferred. Restore verifies exact locked packages from still-available
sources or the destination cache for supported dependency clients: Claude Code,
Codex, OpenCode, Cursor and Gemini CLI.
Authentication, client settings, secrets and plugin data need separate setup or backup. Transferring the files does not guarantee offline installation. See Restore on another machine for confirmation prompts, supported restore behavior and recovery.
Install Agent Skills with agentplugins skills add
from a GitHub repository or from an HTTPS publisher that serves a .well-known
Skills index. Publisher installs take an exact --skill name. The local browser
picker and saved dependencies are included.
Install for one project
SOURCE ONLY / UNRELEASED: select project scope explicitly in a compatible source build:
agentplugins add owner/repo --alias review --target cursor \
--scope project --project-root /path/to/project
agentplugins install --locked --scope project --project-root /path/to/project
agentplugins skills add owner/repository --ui \
--scope project --project-root /path/to/project
The portable uap.toml and uap.lock live in that project. Local receipts,
recovery state and retained data live under .uap/; the user's environment
remains separate. The default scope is still user, and project scope without
--project-root uses the current directory. Native capabilities and activation
requirements depend on the client. See Project installation
for supported-target reporting, update/repair/removal and portable restore.
The current project support table records each
client's component paths and activation limits; user-scope compatibility below
is a separate contract.
VS Code project scope prepares .github/skills and owned .vscode/mcp.json
servers for one local folder. Its Prepared result leaves workspace/server trust
and runtime selection to the client. See VS Code project preparation.
Any Agent Plugins 1.0 package
Install from the registry, a local folder, or a GitHub repository pinned to an
exact commit. plugin.json is the sole supported authoring manifest.
The installed package is standard-first:
plugin.json
├── skills/ optional reusable instructions
├── mcp.json optional MCP servers
└── hooks/ optional client-specific extension
Skills and MCP are portable components. Hooks and other client-specific extensions depend on the target client; they do not imply portable behavior.
You can also install a local package or a pinned GitHub package without adding it to the registry. Direct-install examples are collected near the end of this README.
SOURCE ONLY / UNRELEASED: direct source support now includes arbitrary HTTPS
and SSH Git, GitHub/GitLab/Azure browser URLs, and complete HTTPS/local
ZIP/TAR/TAR.GZ/TGZ packages. See Direct complete package sources
for syntax, native credentials, immutable archive byte pins and supported limits.
These complete-package agentplugins add transports do not import standalone
skill files as plugins. The separate source-build agentplugins skills add
workflow above installs raw Agent Skills through deterministic normalization.
Other manifest formats are ignored and cannot override plugin.json.
Supported clients
The CLI has adapters for:
Grok Build, Kimi Code, and Qoder accept automatic user-scope installation of portable skills and MCP servers through agentplugins add ./my-plugin --target grok,kimi,qoder. Grok Build and Kimi Code support stdio/HTTP MCP; the Qoder adapter currently supports stdio MCP. The CLI verifies native registration and supports update and removal. A running client still needs a reload, and MCP authentication/connectivity is verified in that client.
Qoder also preserves target-native hooks/**, commands/**, and bundled bin/** files for native discovery. Its Go SDK supports SessionStart, UserPromptSubmit, PostToolUse, and Stop. These native files are separate from portable hook and command conversion; portable agents, hooks, commands, and LSP components remain outside these three adapters' supported scope.
Compatibility depends on the package and client. The CLI tells you what is installed, prepared, or still needs activation or sign-in.
See the client compatibility evidence for tested client versions, platforms, and release-specific limitations.
npx universal-agent-plugins add context7 --target chatgpt,kiro
- ChatGPT: the CLI guides you to create a Developer Mode app for
https://mcp.context7.com/mcpwith No authentication, accepts the resultingasdk_app_...ID, and prepares a personal marketplace package. You still install that package and select Context7 in a new chat. When other clients are selected too, the CLI installs them first and reports ChatGPT as a separate setup step instead of cancelling the whole batch. - Kiro: the CLI installs its owned skills and MCP entry while preserving unrelated configuration. On macOS and Windows it reports the remaining Kiro OAuth/restart step instead of claiming automatic runtime verification.
The same retained installation supports repeat add, update, repair, remove, and reinstall. ChatGPT availability still depends on an account or workspace where Developer Mode and custom apps are enabled.
See the real Context7 ChatGPT and Kiro E2E evidence.
A schema pass means that the package is well-formed; it does not prove runtime, OAuth, or activation in every client.
For Codex, declared MCP SSE is unsupported; stdio and Streamable HTTP retain their existing adapter support. Valid SSE components are excluded from Codex delivery without invalidating the package. See transport evidence and lifecycle behavior.
Already built with UAP SDK
One SDK, many agents (reusable installer, typed hooks and more)
Build the next one yourself: how to build with the SDK.
Find and verify plugins
Use agentplugins search or browse the registry to find plugins.
Reviewed registry entries and unreviewed discovery results are labelled separately.
search combines the reviewed Registry Directory with a signed public Discovery
Index containing conformant package paths. Discovery records are
unreviewed metadata, not endorsements. They install only through a
publisher-qualified exact-SHA selector and are validated again before mutation.
The registry is optional for direct installs, but it makes reviewed short names, provenance, compatibility notes, and safe discovery convenient:
Browse the registry · Submit a package
Lifecycle safety
The CLI checks packages before installation and changes only files it manages.
Use --dry-run to preview changes. OAuth and consent stay under your control.
No install telemetry is sent.
Exact removal, verified end to end. remove deletes exactly the files the
CLI added and nothing else: every removed path is checked against installer
ownership, and neighboring files that belong to you or to other tools are left
byte-identical. If a foreign tool has taken over or modified one of the CLI's
managed entries, the CLI refuses to touch it and stops with a clear instruction
instead of overwriting or adopting someone else's configuration. This is
enforced fail-closed, not best-effort: end-to-end tests install and remove a
plugin in every supported agent and record the exact removal inventory
(before/after file digests) as artifacts — across all 13 clients, including
real Claude Code, Codex, and OpenCode runs on macOS and Windows CI runners.
Details and the evidence are in docs/REMOVAL_SAFETY.md.
Before changing a client, the CLI validates the source and preflights every
selected target. --dry-run prints the same plan without writing. Managed files
and state are committed together; failures roll back what ownership proves safe
or stop with a repair command. Preview does not activate a plugin or sign you in.
Repair checks managed files; it does not prove external OAuth or runtime behavior.
The CLI is an independent community project. It is not affiliated with OpenAI, Agent Plugins, or the vendors shown above.
More commands
For normal interactive use, omit --target and choose agents in the prompt.
Use --target in scripts, CI, or whenever you want to name clients explicitly.
# Find and inspect plugins
agentplugins search docs
agentplugins info context7
# Install in specific agents
agentplugins add context7 --target codex,claude
# Manage an installed plugin
agentplugins update context7 --target codex,claude
agentplugins repair context7 --target codex,claude
agentplugins remove context7 --target codex,claude
agentplugins outdated --all
agentplugins update --all
agentplugins doctor
# Install a local Agent Plugins 1.0 package
agentplugins validate ./my-plugin
agentplugins add ./my-plugin
# Install the only Agent Plugins package found at an exact commit
agentplugins add \
owner/repository@0123456789abcdef0123456789abcdef01234567
# Choose a package explicitly when a repository contains several
agentplugins add \
owner/repository@0123456789abcdef0123456789abcdef01234567//path/to/plugin
Remote installs require a full 40-character commit SHA. Branches, tags, and
abbreviated SHAs are rejected. When no package path is given, the CLI uses a
valid root plugin.json or auto-selects the only valid nested package that has
mcp.json or skills/. If several packages match, it lists them and asks for
an explicit //path. Repositories with more than 16 possible packages require
an explicit path before candidate packages are fetched. The selected canonical
path and package digest are stored for safe replay. Direct full-SHA installations
remain immutable; use switch to move to another exact source. repair reapplies
the recorded source, and remove changes only files owned by the CLI.
Build plugins
Create a portable Agent Plugins 1.0 package manually from a root plugin.json, with
optional skills/ and mcp.json, then validate and install it with the commands
above. The current npm and native installer releases do not include
agentplugins author.
The Build guide covers package structure,
remote MCP, stdio MCP, hybrid packages, and the handoff to installation. Its
authoring commands require a qualified authoring build or a source build.
See the Use / Build quickstart.
Native Agent Plugins releases use the agentplugins-v* tag prefix.
In qualified authoring builds, agentplugins author is the public authoring
entrypoint. These examples require such a build or a current source build;
installing the latest npm or native installer does not provide them.
Create and check a Skill package:
agentplugins author init ./my-plugin --template skill --name my-plugin
agentplugins author validate ./my-plugin
agentplugins author inspect ./my-plugin
agentplugins author test ./my-plugin
Current source builds also include JSON maintenance, dependency bootstrap, and
a continuous MCP development loop. Plan mode is read-only, and --write is
required for JSON maintenance changes:
agentplugins author normalize ./my-plugin --document plugin.json
agentplugins author normalize ./my-plugin --document mcp.json --write
agentplugins author import native ./claude.json --from claude \
--output /absolute/path/to/new-plugin --name new-plugin \
--description "Imported Claude MCP package" --write
agentplugins author dev ./my-plugin
Native import reads only the explicit strict-JSON file, skips unsafe or
credential-bearing servers, and publishes only to an absent output. These
commands become supported installation guidance after a signed agentplugins
authoring release passes the public-channel checks.
Authoring validation and project doctor are distinct from installer
agentplugins validate and agentplugins doctor. In qualified/source builds,
MCP execution is available through author test and author dev.
Dependency bootstrap is supported in qualified/source builds. Client projection,
export/bundle, and publication are not exposed by the current authoring CLI.
OAuth and activation inside a supported client still require client-specific
verification.
plugin.json is the only supported authoring manifest. JSON maintenance does
not add a YAML reader, fallback, or migration workflow.
Contributing
Contributor checks:
go test ./...
make vet
- Contributing: CONTRIBUTING.md
- Security policy: SECURITY.md
- Support boundary: docs/SUPPORT.md
- Native CLI installation: docs/NATIVE_INSTALL.md
- Portable commands: docs/UAP_COMMANDS.md
- Client E2E evidence: docs/AGENTPLUGINS_CLIENT_E2E.md
- Registry repository: https://github.com/777genius/universal-agent-plugins-registry
License
Universal Agent Plugins is licensed under the Apache License 2.0. Third-party components retain their original licenses; see NOTICE for attribution.
