LlamaParse

Parse, classify, split and extract data from documents, and build indexes.

Add to Agentman Studio

Opens Agent Studio, where connecting is one click. The connector URL below works in any MCP client.

Paste it into any MCP client. Setup docs

Resources

What the LlamaParse connector does

The LlamaParse MCP server connects LlamaIndex's document-processing platform to Claude, ChatGPT and any MCP-compatible agent. Its 25 live tools parse PDFs and 130+ other formats into markdown, classify and split files, extract structured data against a schema, and build searchable indexes. Sign-in is OAuth, and most calls spend metered credits.

Verified connector

Listed by Anthropic as a partner connector in its Connectors Directory.

Connection checked by Agentman on .

Anthropic states this reflects the level of review a connector received, not a security audit.

LlamaParse tools (25)

  • getUploadUrlActs

    Get a pre-signed URL to upload a file to the LlamaParse S3 storage

  • uploadFileByUrlActs

    Upload a file to LLamaParse S3 storage providing a URL to download the file data. On upload completion, the file will be sent to LlamaParse S3 storage, so that it can be used for downstream processing tasks like parsing, classification or splitting.

  • getUserProjectsRead-only

    List the projects available to the user, with their names, so you can pass the right projectId to other tools. Use this whenever a tool needs a projectId and the correct project is not already known — pick by name, and ask the user if the name is ambiguous.

  • parseFileActs

    Parse a file providing its file ID, retrieving markdown or plain text content of the file. Use with file IDs obtained with the getUploadUrl/uploadFileByUrl tool or that the user provided

  • classifyFileActs

    Classify a file (based on specific categories) providing its file ID. Use with file IDs obtained with the getUploadUrl/uploadFileByUrl tool or that the user provided

  • splitFileActs

    Split a file into category-based segments providing its file ID. Use with file IDs obtained with the getUploadUrl/uploadFileByUrl tool or that the user provided

  • searchSchemaTemplatesRead-only

    Search the built-in library of starter extraction schemas (invoice, contract, resume, 10-K, patient intake and more) by keyword or category. Returns matching templates with their top-level field names, but not the full JSON Schema — call `getSchemaTemplate` for that. Prefer this over `generateExtractionConfig` when the user wants a common document type: it is instant and needs no sample file.

  • getSchemaTemplateRead-only

    Retrieve the full JSON Schema for one starter extraction template, by id, as returned by `searchSchemaTemplates`. Pass the schema to `createExtractionConfigFromSchema` — edit it first if the user needs extra or fewer fields.

  • createExtractionConfigFromSchemaActs

    Create an extraction configuration from a JSON Schema you already have, and return its configuration id for use with `extractFile`. Supply either a `templateId` from `searchSchemaTemplates` (the server loads the schema for you) or an explicit `dataSchema` — a schema you wrote, or a template schema you edited. Unlike `generateExtractionConfig`, this needs no sample file and involves no LLM step.

  • generateExtractionConfigActs

    Generate the configuration to extract structured data from a specific file using the Extract service from the LlamaParse Platform. Provide a prompt describing what the schema of the extracted data, the ID of the file to extract and, optionally, a project ID.

  • extractFileActs

    Extract structured data from a file based on the configuration created with the `generateExtractionConfig` tool. Returns the extracted structured data.

  • listIndexesRead-only

    List all the available indexes on the LlamaParse Platform. Indexes are vector-indexed directories with the possibility of searching/reading/grepping files and performing retrieval.

  • findFilesInIndexRead-only

    Search files within an index. Optionally provide the file name to filter for or a substring that should be contained in the file name

  • readFileFromIndexRead-only

    Read the content of a file from an index, providing its file ID and, optionally, an offset and a maximum length (in characters) to read.

  • grepFileFromIndexRead-only

    Grep the content of a file from an index, providing its file ID, the pattern to grep for and, optionally, a number of context characters and a maximum number of grep matches to retrieve

  • retrieveFromIndexRead-only

    Perform hybrid search on the index, providing a query and, optionally, the top K documents to retrieve and the top N documents to rerank

  • createDirectoryActs

    Create a directory (folder) to hold source documents. A directory is what an index is built over: upload files, add them to a directory with addFilesToDirectory, then call createIndex on it.

  • listDirectoriesRead-only

    List the directories in a project. Returns one page at a time — pass the returned nextPageToken to fetch more. By default only user directories are listed; indexes create their own internal output directories, which are not valid sources for a new index.

  • listDirectoryRead-only

    List the files inside a directory, along with the directory itself. Returns one page at a time — pass the returned nextPageToken to fetch more. Use directoryFileId, not fileId, when referring to a file in later calls.

  • addFilesToDirectoryActs

    Add already-uploaded files to a directory, so they can be indexed. Takes file IDs from getUploadUrl or uploadFileByUrl. Files are added one by one, so the response reports which succeeded and which failed — retry only the failed ones. If the directory already backs an index, call syncIndex afterwards to pull the new files in.

  • createIndexActs

    Create an index over a directory, making its documents searchable with retrieveFromIndex and the other index tools. The directory must already contain the files you want indexed — upload them and call addFilesToDirectory first. Indexing runs in the background: the returned index is not queryable until getIndexStatus reports it ready.

  • getIndexStatusRead-only

    Check whether an index has finished building. Indexing is asynchronous, so an index created or synced moments ago will not return results yet. Poll this until status is 'ready'; 'failed' means the build did not complete. Querying an index that is not ready looks identical to an index with no matching documents.

  • syncIndexActs

    Re-index a directory, picking up files added or changed since the last run. Indexes do not refresh on their own, so this is the only way an existing index sees new documents. Runs in the background — poll getIndexStatus until status is ready. If a sync is already running, this returns syncStarted=false (the underlying API responds 409 or 429): do not retry syncIndex, call getIndexStatus to check the running sync and wait until status is ready.

  • parseWithLiteParseRead-only

    Parse a PDF file with LiteParse, a fast, in-process parser that does not consume credits from the LlamaParse Platform. The tool needs a file ID obtained with the getUploadUrl/uploadFileByUrl tool or provided by the user. Only works with PDF files.

Show the other 1 tools
  • estimateFileComplexityRead-only

    Estimate the parsing complexity of a PDF file (providing its file ID) using LiteParse. Returns a JSON object mapping each page with the LlamaParse tier it should be parsed with (or if you should use LiteParse), based on the parsing complexity and the need for OCR. Use in combination with parseFile and parseWithLiteParse. The tool needs a file ID obtained with the getUploadUrl/uploadFileByUrl tool or provided by the user. Only works with PDF files.

Read from the server on 2026-08-23, including each tool's own safety annotations.

Limits

  • No application scopes exist. The authorization server at https://login.llamaindex.ai advertises openid, profile, email and offline_access — identity claims only — and the RFC 9728 resource descriptor declares no scopes_supported at all. No boundary falls between reading an index and spending credits on a 45-per-page parse. The product-scoped endpoints are the only available substitute, and they are a client configuration choice, not a grant.
  • The tool list is readable without authentication. Our anonymous handshake returned all 25 names, descriptions and schemas. Tool *calls* remain gated, so this is metadata disclosure rather than an open server — but anyone can enumerate the surface, and the descriptions name internal concepts such as configuration IDs and project IDs.
  • No tool can read a credit balance or a job's cost. Metering is central to this product and entirely absent from the tool surface. Budget control has to live outside the agent.
  • No idempotency key on the spending tools. parseFile, extractFile, classifyFile and splitFile declare idempotentHint: false. A retried timeout bills twice.
  • do_not_cache is not exposed. The documented control for keeping a sensitive document out of the 48-hour cache has no MCP equivalent.
  • Nothing can delete. No tool removes a file, directory or index. Storage limits are 10,000 files / 10 GB on Free through 100,000 / 100 GB on Pro, and hitting one returns 402 on new retained uploads. Cleanup is a dashboard task.
  • Rate limits are strict on Free. LlamaIndex documents 20 requests per minute for free-tier organisations, against per-endpoint limits of 50 QPS for file upload and 40 QPS for classify on paid plans, with 429 on exceed.
  • Hard service limits apply. 512 MB maximum file size platform-wide, 100 MB and 500 pages for extraction, 64 KB of text and 35 images extracted per page, and 1–50 categories per split job.
  • We did not verify the product-scoped servers. The /parse/mcp, /extract/{configId}/mcp and sibling endpoints are documented by LlamaIndex; we probed only the unified endpoint.
  • The vendor's "open source" claim did not resolve. LlamaIndex's MCP page states the server is open source. The two plausible repositories, run-llama/llamacloud-mcp and run-llama/mcp-server-llamacloud, are both archived, and the first is a local stdio predecessor with a different design — user-configured indexes and extract agents passed as CLI flags — whose README redirects readers to the docs. We found no public source for the hosted server, so we could not verify its scope enforcement by reading it.

Frequently asked questions

Does the LlamaParse connector upload my documents to LlamaIndex servers?

Yes, and that is the point of it. Every processing tool takes a file ID that only exists after a file reaches LlamaCloud S3 storage, through getUploadUrl or uploadFileByUrl. LlamaIndex states files are cached 48 hours then permanently deleted, and are never used for model training. A parsed contract or medical record leaves your machine.

Can I stop LlamaParse from caching a sensitive document?

Not from this connector. LlamaIndex documents do_not_cache=True as the way to skip the 48-hour cache when submitting a parse or extract job. No tool on the MCP server accepts that parameter — we read all 25 live schemas on 2026-08-23 and found no cache, retention or expiry field. The control exists in the REST API only.

Does calling a LlamaParse tool cost money?

Most of them do. LlamaIndex prices every feature in credits at $1.25 per 1,000, billed per page. Parsing runs 1 credit per page on Fast up to 45 on Agentic Plus, and older v1 modes reach 150. Only parseWithLiteParse is documented as consuming no credits. A single agent call over a long PDF can spend real money.

Can the agent check the credit balance before spending it?

No. We scanned all 25 live tool names for credit, balance, usage, quota and billing on 2026-08-23 and none matched. LlamaIndex exposes usage through the dashboard and a REST endpoint, and per-job credits through expand=usage, but neither is reachable as a tool. The agent can spend without any tool that could check the cost first.

What happens when LlamaParse credits run out?

On paid plans, nothing stops. LlamaIndex states pay-as-you-go is enabled by default on Starter and Pro, so exhausted credits keep billing at $1.25 per 1,000 rather than halting. Only the Free plan hard-stops, returning 402. An exhausted balance that tops itself up is a purchase, not a brake.

How many tools does the LlamaParse MCP server actually have?

Twenty-five, which is three more than either published list. Anthropic's directory and LlamaIndex's own documentation both name exactly the same 22 tools. Our anonymous tools/list on 2026-08-23 returned searchSchemaTemplates, getSchemaTemplate and createExtractionConfigFromSchema in addition. The live server is ahead of the vendor's own page.

Is the LlamaParse MCP server open to unauthenticated callers?

Its tool list is; its tools are not. We completed an anonymous initialize and tools/list against the endpoint on 2026-08-23 with no credentials and read all 25 schemas. The server's RFC 9728 descriptor still requires OAuth, and LlamaIndex states authentication is prompted on first tool use, so nobody can spend your credits without your token.

What OAuth permissions does LlamaParse request?

Only identity claims. LlamaIndex's authorization server advertised openid, profile, email and offline_access on 2026-08-23, and the resource descriptor declares no scopes_supported at all. Nothing at the grant separates reading an index from spending credits on a parse, so consent is all-or-nothing across the platform.

Can the LlamaParse connector delete my files or indexes?

No. No tool among the 25 carries a delete, remove, drop or archive verb, and none is marked destructiveHint true by the server itself. The writes create directories, indexes and configurations, and add files. To free storage you delete files in the dashboard, which is outside this connector's reach.

Sources

More Productivity connectors

  • 6sense

    Ask an AI agent about 6sense account insights, buying signals and keyword trends.

  • Adobe Workfront

    Find, create, update and delete Workfront work items from an AI agent.

Browse all MCP connectorsClaude and ChatGPT setup guideBrowse AI skills

Product names and logos belong to their owners. Agentman is not affiliated with a connector's maker unless the page says the connector is built and run by Agentman.

Use it in an agent

Put LlamaParse to work.

Connect LlamaParse once and your agents call these tools on their own: on a schedule, in a workflow, with nobody at the keyboard.

Open in Agentman Studio
Call (650) 285-1019Our AI receptionist answers.