Plugins
A plugin is a namespaced bundle of customizations that ships, installs, and
turns on as one unit. It can contain skills, agents, rules, MCP servers and
hooks. When a plugin is on, everything inside it is active; disabling it takes
the whole bundle down at once.
Your job is either to change a plugin's lifecycle state or to get new
functionality into the right plugin. Step 0 and Step 1 are common to both; Step
1 routes you to Step 2A or Step 2B.
For the underlying customization system — discovery roots, loading priority,
what each customization type is for — see the customizations guide skill. This
skill does not restate that material.
Quick Start
After writing /tmp/lsrpc.sh in Step 0, prefix every lsrpc call with
source /tmp/lsrpc.sh && and key every mutation on <dir> (the install
directory name (.path | split("/") | last) from Step 0, never plugin.json's
name):
<!-- mdformat off -->
| Operation | Command / Action | Key Guardrail |
|---|
| Inventory installed plugins (Step 0) | source /tmp/lsrpc.sh && lsrpc GetAllPlugins '{"workspaceUris":["file:///path/to/workspace"]}' | jq '[.plugins[] | {name, dir: (.path | split("/") | last), disabled: (.disabled // false), isGlobal, skills: [.skills[]?.name], agents: [.agents[]?.name], rules: [.rules[]?.absolutePath], hooks: ((.hooksJson // "") | length > 0), mcp: [.mcpServers[]?.name]}]' | Never read GetAllPlugins unfiltered (inlines full SKILL.md text); always pipe through jq and use dir for mutations ('{}' if no workspace is active). |
| Enable / Disable plugin (Step 2A) | source /tmp/lsrpc.sh && lsrpc JetboxWriteState '{"userConfig":{"plugins":{"<dir>":{"enabled":true}}}}' (false to disable) | Takes effect live (no restart). Never hand-edit ~/.gemini/config/config.json. |
| Uninstall user plugin (Step 2A) | source /tmp/lsrpc.sh && lsrpc DeletePlugin '{"pluginId":"<dir>"}' | Refuse if .path contains /builtin/plugins/ (offer to disable instead). If currently disabled, set "enabled":true via JetboxWriteState before deleting. |
| Install Marketplace plugin (Step 2A) | source /tmp/lsrpc.sh && lsrpc InstallCustomization '{"id":"<marketplace-id>"}' | Use <marketplace-id> from the Customizations page or user (never guess from name; only Marketplace plugins can be installed). Requires a restart. |
| Author / consolidate plugin (Steps 2B–4) | Create ~/.gemini/config/plugins/<name>/plugin.json (or in source repo if published) + used subdirectories (skills/, agents/, rules/AGENTS.md, mcp_config.json, hooks.json, assets/logo.svg) | Check Step 0 inventory for name collisions first; re-run Step 0 inventory, tell the user to restart, and link as [<name>](customization://<dir>) (never plugin://). |
<!-- mdformat on -->
Step 0: Preflight and inventory
Every path needs to know what is already installed. Talk to the language server
over its local Connect endpoint. The address and CSRF token are normally already
in your shell environment; you do not need to hunt for a port file.
: "${ANTIGRAVITY_LS_ADDRESS:=localhost:5387}"
curl -sS "http://${ANTIGRAVITY_LS_ADDRESS}/healthz"
Write the helper to a file once, then source it in every call below. Each
command you run is a fresh shell, so a function defined in one call is gone by
the next; sourcing a file is what makes it reusable.
cat > /tmp/lsrpc.sh <<'EOF'
: "${ANTIGRAVITY_LS_ADDRESS:=localhost:5387}"
lsrpc() { # usage: lsrpc <MethodName> <json-body>
curl -sS -X POST \
"http://${ANTIGRAVITY_LS_ADDRESS}/exa.language_server_pb.LanguageServerService/$1" \
-H "Content-Type: application/json" \
-H "x-codeium-csrf-token: ${ANTIGRAVITY_CSRF_TOKEN}" \
-d "$2"
}
EOF
Now list the installed plugins, passing active workspace URIs in
workspaceUris so workspace-scoped plugins are discovered (or '{}' when no
workspace is active). Prefix every lsrpc snippet in this skill with
source /tmp/lsrpc.sh &&:
source /tmp/lsrpc.sh && lsrpc GetAllPlugins '{"workspaceUris":["file:///path/to/workspace"]}' | jq '[.plugins[] | {
name,
dir: (.path | split("/") | last),
disabled: (.disabled // false),
isGlobal,
skills: [.skills[]?.name],
agents: [.agents[]?.name],
rules: [.rules[]?.absolutePath],
hooks: ((.hooksJson // "") | length > 0),
mcp: [.mcpServers[]?.name]
}]'
Never read GetAllPlugins unfiltered. The response inlines the full
text of every bundled SKILL.md. A single plugin can exceed 16KB and a
populated config will flood your context. Always pipe it through jq.
dir is the identifier that matters. It is the install directory name — the
last path segment — and it is what every mutation below is keyed on. It is
frequently not the same as name from plugin.json: a plugin installed
from the Marketplace lands in a directory named after a signed 64-bit hash, so a
plugin displayed as data-tools may live in a directory called
-1870732469773246571. Using name where dir is expected fails silently or
builds a garbage path.
Step 1: Decide the kind
- The user names an existing plugin and a lifecycle verb — "turn off X",
"install Y", "get rid of Z" -> Step 2A.
- The user wants new capability, or wants existing loose customizations tidied
up — "I want a Strava thing", "package these skills together" -> Step
2B.
Step 2A: Lifecycle operations
Resolve the target to its dir from the Step 0 inventory first. If the user's
words match several plugins, ask which one rather than guessing.
Enable or disable. This is the same write the Customizations page performs,
and it takes effect live — no restart.
# Enable:
source /tmp/lsrpc.sh && lsrpc JetboxWriteState '{"userConfig":{"plugins":{"<dir>":{"enabled":true}}}}'
# Disable:
source /tmp/lsrpc.sh && lsrpc JetboxWriteState '{"userConfig":{"plugins":{"<dir>":{"enabled":false}}}}'
The write deep-merges, so send only the one plugin. Never hand-edit
~/.gemini/config/config.json to do this: writes that bypass the language
server miss the live reload and are only picked up by a slow backstop poll.
Uninstall.
source /tmp/lsrpc.sh && lsrpc DeletePlugin '{"pluginId":"<dir>"}'
Check both of these before you run it:
- Built-in plugins cannot be uninstalled. A plugin whose path contains
/builtin/plugins/ ships with the product. Refuse, and offer to disable it
instead.
- Uninstall leaves the enablement entry behind. Nothing clears
plugins.<dir> from config.json. If the plugin was disabled and is later
reinstalled into the same directory name, it comes back disabled for no
visible reason. Before deleting a disabled plugin, set it back to enabled so
the stale entry is harmless, and tell the user you did.
Install. For a plugin published to the Agent Marketplace:
source /tmp/lsrpc.sh && lsrpc InstallCustomization '{"id":"<marketplace-id>"}'
You cannot derive this ID from the plugin's name. Take it from the Marketplace
tab of the Customizations page, or ask the user for it. If all you have is a
name, say so rather than guessing an ID.
A source not found error means one of two things: the ID is wrong, or the
Marketplace catalog isn't launched for this user. The catalog case has a tell —
the ID comes back empty, so the message reads source for customization not found with nothing between the words. Re-check the ID first.
Only install plugins that have a Marketplace entry. If the plugin has none, or
the catalog isn't launched for this user, do not try to install it another way;
tell the user it isn't available to install.
A newly installed plugin directory is only discovered on startup. Tell the
user to restart before expecting its skills to appear.
Lifecycle operations end here. Skip Step 3, which covers authoring only, and go
to Step 4.
Step 2B: Decide where the functionality belongs
A plugin is a bundle, so the first question is never "what do I build" but
"where does it go". Take the inventory from Step 0 and see whether the request
fits an existing plugin. Plugins tend to cohere around one of three shapes:
- A data provider or external service. Everything that talks to one system
— Strava, GitHub, a particular API. Its MCP server, the skill that drives
it, and the rules about its quirks belong together.
- An artifact or file type. Spreadsheet handling, proto editing, log
analysis. Grouped by what the user is operating on.
- A role or recurring job. Data analysis, running a small business,
reviewing documents, writing code. Broader, and the right home when a
capability spans several providers in service of one repeated task.
Recommend and proceed. State in one line which plugin you picked and why,
then keep going. Use ask_question only when two existing plugins are
genuinely close, or when nothing fits and the boundary for a new one is unclear.
Do not stall on a decision the user can trivially reverse.
If nothing fits, create a new plugin at ~/.gemini/config/plugins/<name>/.
Before settling on a name, check it against the Step 0 inventory: names share
one global namespace, and when two plugins claim the same name the first
discovered wins and the other is dropped silently.
If the plugin is meant to be shared or published, create it in the source
repository it will be published from instead of the user's config directory.
Consolidating loose customizations
When the user asks to tidy up standalone customizations, move the directories: a
skill at ~/.gemini/config/skills/foo/ becomes
~/.gemini/config/plugins/<name>/skills/foo/, and a rule becomes
rules/AGENTS.md inside the plugin.
MCP servers need a warning first. Moving an entry out of the user's
mcp_config.json and into the plugin's own mcp_config.json ties its lifecycle
to the plugin: it is assembled only while the plugin is enabled, so disabling
the plugin stops its tools being offered along with everything else in the
bundle. Control moves with it too, since the server is no longer an entry in the
user's own config and is managed through the plugin rather than on its own. Tell
the user this before you move the entry.
Step 3: Author the bundle
Create only the subdirectories the plugin actually uses:
~/.gemini/config/plugins/<name>/
├── plugin.json # required
├── README.md # user-facing docs; never read by the agent
├── mcp_config.json # optional, same format as the user's own
├── hooks.json # optional
├── assets/logo.svg # optional
├── skills/<skill>/SKILL.md
├── agents/<agent>.md
└── rules/AGENTS.md
Write the inner customizations yourself — the SKILL.md, the agent markdown,
the rules, the MCP config. Do not hand that back to the user as a follow-up.
Rules in a plugin's rules/AGENTS.md are plain markdown with no frontmatter and
are always on while the plugin is enabled.
plugin.json is parsed as JSONC, so comments and trailing commas are allowed:
{
"name": "strava",
"description": "Strava integration: pull activity data, summarize training load, and draft workout notes.",
"logo": "assets/logo.svg",
"suggestedPrompts": [
"Summarize this week's workout mileage and heart rate zones",
"Compare pacing and cadence across my last 4 long runs",
"Draft an annotated Strava description for my latest activity"
]
}
| Field | Type | Notes |
|---|
name | string | The label users see, and the global unique |
: : : key — two plugins sharing a name collide : | | |
| : : : and one is dropped. Not the install : | | |
| : : : identifier; that is the directory. Lowercase : | | |
| : : : kebab-case. Defaults to the directory name : | | |
| : : : if omitted. : | | |
description | string | One or two sentences saying what the plugin |
| : : : does and who it is for. This is what : | | |
| : : : Marketplace search matches. : | | |
logo | string | Optional. A path relative to the plugin |
: : : directory, such as assets/logo.svg, ending : | | |
: : : in .png, .svg, .jpg, .jpeg or : | | |
: : : .webp. URLs and absolute paths are : | | |
| : : : rejected silently — a bad value renders no : | | |
| : : : logo rather than erroring. : | | |
suggestedPrompts | string[] | Optional. Up to 3 starter prompts for card |
| : : : and detail views. Exactly 3 prompts are : | | |
| : : : recommended for UI card layouts. Empty : | | |
| : : : strings are skipped and additional prompts : | | |
| : : : beyond 3 are ignored. : | | |
displayName | string | Optional. Human-readable presentation title |
| : : : shown in UI headers, cards, and slash : | | |
| : : : command menus. : | | |
version | string | Optional. Semantic version string. |
author | string / | Optional. Developer or team name string, or |
: : object : an object with name and optional url : | | |
: : : (https://). : | | |
homepage | string | Optional. External https:// website or |
| : : : documentation URL. : | | |
repository | string / | Optional. Source repository https:// URL |
: : object : (or {"type": "git", "url": "https://..."}). : | | |
disabled | boolean | Optional. Ships the plugin turned off. A |
| : : : choice the user has already made wins over : | | |
| : : : this field. : | | |
Do not add any other top-level field. Unknown top-level fields are
silently discarded by the loader — they neither take effect nor error out.
Offer to generate a logo with the generate_image tool. Ask for a square asset
of at least 128x128 with a transparent background, high contrast so it reads on
both themes, and roughly 10% empty margin baked in — the frame adds no padding
of its own and center-crops anything non-square.
Step 4: Verify and hand back
Re-run the Step 0 inventory and confirm the plugin appears with the components
you expect. Then tell the user:
- What changed, and which plugin now owns it.
- Whether a restart is required. Creating or installing a new plugin
directory needs one; enabling, disabling and uninstalling do not.
- Where to manage it, as a link:
[<name>](customization://<dir>) opens the
plugin's own page, which lists what it provides and can turn it on and off.
<dir> is the install directory from Step 0, never the manifest name. A
plugin directory created in this session is not discovered until a restart, and
until then its page reports nothing installed under that id, so give the link
and the restart together.
Do not emit plugin:// or plugins:// links; they are not handled and render
as dead text.
Constraints & Tips
- Key every mutation on the install directory name, never
plugin.json's
name. This is the single most common way to break a plugin operation.
- Prefer the language server RPCs over editing files.
config.json is
written atomically at mode 0600 and holds sibling state — userSettings,
permission grants. A careless rewrite clobbers them.
- One coherent purpose per plugin. Bundling unrelated capabilities forces
the user into an all-or-nothing toggle. If a request does not fit the
plugin's description, that is the signal to create a new one.
- MCP server names are global too. They collide across the user's config
and every installed plugin, so prefix a plugin's servers with the plugin
name rather than shipping a generic
search or db.
- Keep
description honest and current. It is what the user reads before
installing and what search matches. Update it when the bundle's contents
change.