LlamaParse Agent Plugins

Four document-focused plugins from LlamaIndex, packaged as a marketplace for both Claude and Codex. Parse documents locally with LiteParse, or connect to the LlamaParse Platform for cloud parsing and agentic retrieval over Index v2 knowledge bases.

Which MCP plugin? Region follows your LlamaCloud account, not a per-session choice. Accounts on cloud.llamaindex.ai need llamaparse-mcp; accounts on cloud.eu.llamaindex.ai need llamaparse-mcp-eu. A token issued in one region is rejected by the other.

Install

Add the marketplace:

/plugin marketplace add run-llama/llamaparse-agent-plugins

Then install whichever plugins you want:

/plugin install liteparse@llamaparse-marketplace
/plugin install llamaparse@llamaparse-marketplace
/plugin install llamaparse-mcp@llamaparse-marketplace
/plugin install llamaparse-mcp-eu@llamaparse-marketplace

Skills are namespaced by plugin, so invoke them as /<plugin>:<skill>:

/liteparse:liteparse
/llamaparse:llamaparse
/llamaparse-mcp:llamaparse-mcp   # parse / split / classify
/llamaparse-mcp:llamacloud-index # Index v2 agentic retrieval

The llamaparse-mcp-eu plugin exposes the same skills under its own namespace (/llamaparse-mcp-eu:llamaparse-mcp, and so on).

If the install summary says Run /reload-plugins to activate, run it.

Add the marketplace:

codex plugin marketplace add run-llama/llamaparse-agent-plugins

Then start or reload Codex, open plugin management, find the marketplace llamaparse-marketplace (display name: LlamaParse Codex MarketPlace), and install any of liteparse, llamaparse, llamaparse-mcp, or llamaparse-mcp-eu.

Confirm availability in a Codex thread — skills appear as liteparse:liteparse, llamaparse-mcp:llamaparse-mcp, llamaparse-mcp:llamacloud-index, and so on.

The MCP plugins declare authentication: ON_USE in the Codex marketplace, so Codex prompts on the first tool call rather than at install time.

Authentication

• liteparse — none. Runs fully locally.
• llamaparse — needs a LlamaParse API key in LLAMA_CLOUD_API_KEY.
• llamaparse-mcp / llamaparse-mcp-eu — OAuth, no API key. The flow runs in your client on first tool use.

Verify it works

1. Request an upload URL with getUploadUrl.
2. Upload a file (for example test.txt) to the returned pre-signed endpoint.
3. Call parseFile with the returned fileId.

Expected result: parsed text content is returned.

1. Call listIndexes to confirm at least one index is visible.
2. Call retrieveFromIndex with a natural-language question against one of the returned indexes.

Expected result: relevant passages from your indexed documents are returned.

An empty listIndexes result still means auth succeeded — you just have no indexes yet.

Privacy Policy & Terms

liteparse runs entirely locally and transmits no data. The llamaparse, llamaparse-mcp, and llamaparse-mcp-eu plugins send the documents you ask them to process to the LlamaParse Platform for parsing, indexing, and retrieval.

• Privacy Notice
• Terms of Service
• Support: [email protected]

Repository layout

.claude-plugin/marketplace.json    marketplace definition consumed by Claude
.agents/plugins/marketplace.json   marketplace definition consumed by Codex
plugins/<name>/
  .claude-plugin/plugin.json       Claude manifest
  .codex-plugin/plugin.json        Codex manifest
  .mcp.json                        MCP server config (MCP plugins only)
  skills/<skill>/SKILL.md          skills, one directory each
  assets/                          icon and logo

Each plugin carries a manifest per agent. The two must agree on name and version — the validator enforces this.

Development

Validate

CI runs this on every pull request, and it is worth running before you push:

python3 scripts/validate-marketplace.py

It checks that both marketplace manifests list the same plugins, that every registered plugin exists on disk (and vice versa), that each plugin.json name matches its directory, that the Claude and Codex versions of a plugin agree, that .mcp.json and the inline mcpServers block declare the same server, that no two plugins claim the same MCP server key, and that every skill directory has a SKILL.md.

For Claude specifically, also run the first-party validator — the plugin directory's review pipeline runs the same check:

claude plugin validate .
claude plugin validate ./plugins/<name>

Bump version

python3 scripts/bump-version.py <claude|codex> <plugin> --bump <patch|minor|major>
# if you have uv
./scripts/bump-version.py <claude|codex> <plugin> --bump <patch|minor|major>

Bump both agents so the versions stay in sync:

./scripts/bump-version.py codex liteparse --bump minor
./scripts/bump-version.py claude liteparse --bump minor

Keeping skills in sync

Most skills here are copies with an upstream elsewhere. There is no automation, so they drift silently. Check them before a release.

llamacloud-index shares a name with the upstream skill but is a different document: upstream drives the REST API with curl and an API key, while the version here drives the MCP tools over OAuth. They are not interchangeable — treat them as separate skills that happen to share a title.

Two deltas in llamaparse-mcp are intentional and should survive a sync: the Index v2 Retrieval section (which points at the sibling llamacloud-index skill, something the upstream server repo has no notion of), and the cross-region 401 advice (upstream tells the user to repoint their client, which a plugin user cannot do because the URL is fixed in the manifest — here the instruction is to install the sibling regional plugin).

The EU plugin's skills are copies of the NA plugin's with the hostname and region swapped. Change one, change the other.