# Claude Project File Size Limit: The 30MB Upload Ceiling and MCP Storage

The Claude Project file size limit is Anthropic's 30MB per-document ceiling on individual files uploaded to a Project persistent knowledge base, compared to 500MB in standard chats. While project file counts are officially unlimited, cumulative text must fit within the model context window before automated retrieval activates. Teams working with extensive technical libraries avoid context exhaustion by storing documents in an external workspace and querying them on demand via MCP.

Source: https://fast.io/resources/claude-project-file-size-limit/
Author: [Tom Langridge](https://fast.io/authors/tom-langridge/)
Last reviewed: 2026-10-15

## What Is the Claude Project File Size Limit?

Anthropic restricts individual files uploaded to Claude Projects to 30MB each, while direct chat conversations accept up to 20 files at up to 500MB each, according to Anthropic's official file upload documentation. Senders arriving at Claude Projects with dense archives often hit this threshold unexpectedly, expecting the 500MB upload capacity available in direct conversations.

The Claude Project file size limit is Anthropic's 30MB per-document ceiling on individual files uploaded to a Project persistent knowledge base, compared to 500MB in standard chats. While standard chats treat attachments as temporary conversational inputs, Claude Projects treats uploaded files as persistent project knowledge that informs every prompt generated within that workspace. This architectural divergence explains why the limits differ so dramatically.

Many published guides confuse the 500MB chat attachment ceiling with the 30MB project knowledge limit. They fail to explain how document parsing and context window capacity enforce these differing rules. In a standard chat, an uploaded file is processed once for that specific discussion thread. In a Claude Project, every uploaded document must be indexed and prepared for persistent reference across dozens or hundreds of future conversations.

The table below contrasts documented file handling limits across Claude environments based on Anthropic documentation verified in October 2026:

| Environment | Max File Size | File Count Limit | Processing Mode | Context Window & Constraints | Source & Date |
| --- | --- | --- | --- | --- | --- |
| Claude Standard Chat | 500MB per file | Up to 20 files per chat | Ephemeral in-context processing | Scoped to single chat turn | Verified October 2026 (support.claude.com) |
| Claude Projects Knowledge | 30MB per file | Unlimited (context-bounded) | Persistent in-context / automated RAG mode | 200,000 tokens base context (up to 10x with RAG) | Verified October 2026 (support.claude.com) |
| Claude Code Projects | 30MB per file | Unlimited repository files | In-context with local file access | Model context tier (200k to 1M tokens) | Verified October 2026 (docs.claude.com) |
| Claude Direct API Requests | 30MB payload | Batch requests | Stateless model calls | Model context tier (200k to 1M tokens) | Verified October 2026 (docs.claude.com) |
| Fast.io Workspace with MCP | 25 GB to 100 GB by plan | Unlimited files and folders | Dynamic semantic indexing (hybrid search) | Zero context overhead until retrieved via MCP | Verified October 2026 (fast.io) |

A persistent myth in online discussions claims that Claude Projects enforces a hard ceiling of 50 uploaded files. That claim is incorrect. Anthropic's official documentation confirms that the number of files in Claude Projects is unlimited, but the total content must fit within Claude's context window. Early beta users frequently assumed a 50-file cap because typical technical documents and specifications consume roughly 3,000 to 4,000 tokens each, causing a 50-file collection to reach the standard 200,000-token context boundary. If you upload concise markdown notes that average 500 tokens each, Claude accepts dozens or hundreds of files without complaint.

Document formatting also governs ingestion behavior. Documents with 100 pages or fewer retain visual analysis, whereas documents spanning 101 to 1000 pages undergo text-only extraction. Any PDF that exceeds 1000 pages triggers an upload error. Non-PDF formats, including DOCX, CSV, TXT, HTML, and JSON, undergo text extraction only. If a Word document contains embedded architecture diagrams, Claude reads the body text but cannot interpret the images.

## Why Is the Claude Project File Limit Smaller Than Regular Chat?

Understanding why Claude Projects restricts individual files uploaded to Claude Projects to 30MB while direct chat conversations accept up to 20 files at up to 500MB each requires examining how Claude stores, tokenizes, and recalls project context.

When you attach a file to a standard chat, Claude parses the document once, injects its contents into that specific session prompt, and lets the interaction run its course. When that chat concludes or scrolls out of memory, those tokens no longer draw computational resources.

Claude Projects operates under an entirely different operational pattern. Files added to Project Knowledge become permanent reference material. By default, the text extracted from every project file is loaded into the model prefix for every new conversation initiated inside that project. This architectural requirement imposes four major constraints:

1. Prompt Prefix Weight: Restricting individual files uploaded to Claude Projects to 30MB prevents users from attaching documents that exceed the active context window. Even an individual 30MB file uploaded to Claude Projects expands into hundreds of thousands of tokens when parsed as raw text.

2. Prompt Caching Economics: Anthropic uses prompt caching to keep project conversations responsive and cost-effective. By keeping the static prefix identical across conversation turns, Claude avoids recomputing attention states for repeated documents. Maintaining this cache across multi-tenant infrastructure requires predictable document boundaries. When individual files uploaded to Claude Projects approach the 30MB ceiling, initial text parsing requires substantial memory before prompt caching takes over.

3. Ingestion Worker Bounds: When a document is uploaded, Claude's ingestion pipeline extracts text, checks formatting compliance, strips binary artifacts, and generates token counts. While direct chat conversations accept up to 20 files at up to 500MB each, parsing massive spreadsheets demands extensive processing time. Enforcing the 30MB boundary on Claude Project knowledge protects the persistent ingestion layer from worker timeouts.

4. Context Budget Token Density: The visual capacity meter in the Claude Project sidebar does not measure megabytes stored on disk. Instead, it measures the percentage of Claude's context window consumed by the text extracted from your uploaded files. Different file types consume tokens at vastly different rates:

* Plain Text and Markdown: Standard English prose averages approximately 1 token for every 3.5 to 4 characters, or roughly 750 words per 1,000 tokens. A 20,000-word software manual consumes about 26,600 tokens, fitting comfortably within the 200,000-token context budget.
* Source Code and Configuration: Source code consumes tokens at a much higher density than conversational text. Indentation, brackets, syntax symbols, camelCase identifiers, and package imports produce high token-to-character ratios. A repository export spanning thousands of lines of code easily absorbs most of the available context budget.
* Structured CSV and TSV Spreadsheets: Tabular data carries extreme token overhead. Every cell value, delimiter, quotation mark, and header name generates distinct tokens. Repeating header names and punctuation across thousands of rows causes a modest CSV document to consume the entire project context window.
* PDFs and Scanned Documents: PDF files contain internal font maps, vector coordinates, layout metadata, and repeated headers and footers. The extracted text often carries formatting overhead that inflates token counts beyond the raw word count.

## What Happens When You Exceed the Claude Project Knowledge Ceiling?

When you push Claude Projects beyond its documented limits, you encounter two distinct barriers: the individual file size ceiling and project knowledge context saturation.

The first barrier is mechanical. If you attempt to upload an individual file that exceeds 30MB, Claude rejects the upload immediately. The interface displays an error message indicating that the file is too large, and the file cannot be saved to the project knowledge list.

The second barrier is context saturation. As you add files uploaded to Claude Projects under the 30MB limit, the project capacity meter fills. When project knowledge approaches the context window boundary, two major performance penalties occur:

First, conversational headroom contracts. Because static project files consume token budget upfront, less space remains for ongoing dialogue. In a project where documentation consumes most of the 200,000-token window, every new chat begins with limited space available. That remaining space must accommodate your questions, Claude's generated answers, thinking tokens, and conversation history. Automatic compaction kicks in much earlier in the dialogue, forcing Claude to summarize previous turns and potentially lose track of subtle project requirements.

Second, model attention suffers from needle-in-a-haystack dilution. When a model must attend to vast quantities of static text in every prompt turn, precision can drop. Relevant technical constraints buried inside secondary documents might be overlooked when answering specific implementation questions.

To alleviate capacity saturation, Anthropic provides automated retrieval for projects on paid subscriptions (Pro, Max, Team, and Enterprise). When Claude Project knowledge approaches context window limits, Claude enables RAG mode to expand capacity by up to 10x, according to Anthropic's official RAG documentation.

When automated RAG activates, Claude alters its retrieval approach. Instead of loading every uploaded document into prompt memory simultaneously, Claude indexes the project files and calls an internal project knowledge search tool to fetch relevant passages dynamically.

While automated RAG provides breathing room, it introduces four practical limitations for engineering and knowledge management teams:

1. Rigid Project Silos: Files uploaded to a Claude Project are permanently restricted to that specific project. If you have three teams working on frontend architecture, backend services, and API documentation, each team must maintain duplicate copies of shared specifications. When a core document changes, every project must be manually updated.

2. The 30MB File Cap Persists: If you possess large server logs or research archives that exceed the 30MB ceiling for individual files uploaded to Claude Projects, Anthropic refuses the upload regardless of your plan tier. Enabling RAG mode increases total project token capacity, but it does not expand the per-document upload boundary.

3. Multi-Turn Retrieval Latency: In RAG mode, Claude must formulate search queries, retrieve excerpts, evaluate their relevance, and synthesize an answer. When project knowledge grows deep and complex, multi-document research questions can introduce noticeable latency compared to direct in-context recall.

4. Bounded Capacity for Large Archives: Even when Claude enables RAG mode to expand capacity by up to 10x, large engineering teams quickly exceed what internal project knowledge can store. Multi-gigabyte code repositories, compliance archives, and corporate document libraries require an external architecture.

## How to Query Files That Exceed Claude Project Limits Using MCP

Teams managing document libraries that exceed Claude's 30MB upload cap or base context window require an architecture that decouples document storage from prompt memory.

Rather than fragmenting files into artificial parts or constantly purging project knowledge, you can place the complete corpus in an external intelligent workspace and connect Claude using the Model Context Protocol (MCP).

In this architectural pattern, your files live inside [Fast.io](https://fast.io). When documents enter a workspace, Intelligence Mode automatically indexes their contents for hybrid search, combining full-text keyword matching, semantic vector retrieval, and structured metadata queries. Instead of forcing Claude to ingest hundreds of megabytes upfront, Fast.io serves as an external memory bank that Claude queries dynamically.

Connecting Claude to Fast.io takes only a few minutes. Fast.io provides a hosted remote MCP server over Streamable HTTP at `https://mcp.fast.io/mcp/tools`. The setup steps for Claude web, desktop, and mobile applications are straightforward:

1. Open Claude and navigate to Customize, then select Connectors.
2. Choose Add Custom Connector and enter `https://mcp.fast.io/mcp/tools` as the connector URL.
3. Select Connect. A secure browser window opens prompting you to sign in to Fast.io via OAuth.
4. On the Review Permissions screen, select whether Claude receives Read Only or Read & Write access, and choose which organizations and workspaces Claude can reach.
5. In any chat conversation, open the `+` menu under Connectors and enable Fast.io.

For developers using coding agents such as Claude Code, Cursor, or VS Code Copilot agent mode, Fast.io provides a dedicated code-oriented MCP endpoint at `https://mcp.fast.io/mcp/code`. In Claude Code, you can register the connection from your terminal:

```bash
claude mcp add --transport http fast-io https://mcp.fast.io/mcp/code
```

Run `/mcp` inside Claude Code to complete authorization. Detailed configuration parameters and reference guides are available at [https://mcp.fast.io/docs](https://mcp.fast.io/docs) and [https://mcp.fast.io/skill.md](https://mcp.fast.io/skill.md).

Once connected, Claude interacts with your workspace through a consolidated MCP toolset. When you ask a question regarding your technical library, Claude calls Fast.io's storage search tools in the background. Fast.io executes the search across indexed files, extracts the most relevant passages with precise source attribution, and passes only those matching snippets back to Claude.

Fast.io organization plans provide scalable capacity for engineering teams, as detailed on the [Fast.io pricing page](/pricing/):

| Plan | Monthly Price | Included Storage | Included Workspaces | AI Credits | Max Upload Size |
| :--- | :--- | :--- | :--- | :--- | :--- |
| Starter | $9.99/mo | 250 GB | 5 workspaces | 100,000 credits | 25 GB |
| Business | $49.99/mo | 5 TB | 50 workspaces | 600,000 credits | 50 GB |
| Enterprise | $199.99/mo | 25 TB | 200 workspaces | 3,000,000 credits | 100 GB |

Monthly plans start with a 14-day free trial that requires a credit card. Credits meter AI work. Storage beyond plan allowances is billed at 1.5 cents per GB a month, and extra bandwidth is billed at 4 cents per GB.

This external retrieval model solves the 30MB file ceiling completely:

* File Size Scaling: Fast.io workspace upload limits accommodate large files directly. Maximum upload sizes scale across Starter, Business, and Enterprise plans to accommodate massive archives directly. You can upload large disk images, extensive dataset dumps, and monolithic technical manuals without manual splitting.
* Zero Upfront Context Consumption: Because files remain in external storage until queried, your Claude conversations start with their full context window open. The base prompt remains clean, leaving maximum token budget for complex reasoning, step-by-step problem solving, and detailed responses.
* Current Updates Across Tools: When you update a specification in Fast.io, the indexed knowledge updates for all connected agents. Claude on the web, Claude Code in your terminal, and human teammates all query the exact same current version.

## Architectural Comparison: Claude Projects vs. External Intelligent Workspaces

Evaluating whether to keep documents inside Claude Projects or delegate storage to an external MCP workspace depends on document volume, update cadence, team size, and workflow complexity.

The table below contrasts native Claude Projects against an external Fast.io intelligent workspace:

| Capability | Claude Projects Knowledge | Fast.io Workspace via MCP |
| --- | --- | --- |
| Per-File Upload Ceiling | 30MB per document | 25 GB (Starter), 50 GB (Business), 100 GB (Enterprise) |
| Total Storage Capacity | Bounded by context window (200k tokens base; up to 10x with RAG) | 250 GB (Starter), 5 TB (Business), 25 TB (Enterprise) |
| Cloud Storage Ingestion | Manual file uploads from local disk | Cloud Sync (Google Drive, Dropbox, Box, OneDrive) and URL Import |
| Workspace Scope & Sharing | Siloed within individual Claude Projects | Shared org-owned workspaces accessible by humans and agents |
| Agent Interoperability | Claude only (web, desktop, mobile) | Claude, Claude Code, Cursor, Codex, ChatGPT, and custom agents |
| Version History | No version tracking (overwrites delete prior files) | Per-file version history on every file |
| Audit Trail | No change tracking | Detailed activity log recording uploads, edits, moves, and deletions |
| Structured Extraction | Unstructured text extraction | Metadata Views with typed schemas (dates, numbers, JSON, booleans) |

Five core architectural advantages distinguish the external workspace model:

1. Unified Repository Across Multiple Tools: Storing knowledge in Claude Projects locks that data inside Anthropic's ecosystem. If your team also uses Cursor for software development, ChatGPT for drafting, or internal agent pipelines for automation, each tool requires its own separate data upload. An external workspace acts as a neutral knowledge hub. One version of your technical archive serves every connected agent through standard MCP protocols.

2. Cloud Synchronization Without Local File Transfer: Claude Projects requires you to download files to your local drive and manually drag them into the browser. Fast.io includes Cloud Sync, which connects Google Drive, Dropbox, Box, and OneDrive folders to a workspace, one-way or two-way, on a schedule or on demand. If your documentation repository lives in Google Drive or Box, updates sync into the workspace automatically without manual export steps. One-time URL imports can also pull files directly from external cloud links.

3. Structured Extraction via Metadata Views: While Claude Projects performs plain text extraction, complex business and technical files often contain structured tabular records, dates, and contract terms. Fast.io provides [Metadata Views](/product/document-data-extraction/), which turn documents into queryable, typed schemas. You can define fields in natural language (such as effective dates, counterparties, or release versions), and the extraction engine matches files across the workspace to populate a filterable spreadsheet. Connected agents can query these structured attributes directly via MCP rather than reading full-text files.

4. Versioning and Auditing for Concurrent Work: In Claude Projects, replacing a document overwrites the previous version permanently, offering no rollback mechanism or edit record. Fast.io maintains per-file version history, allowing teams to review changes and restore previous versions. In addition, a detailed activity log tracks all organizational modifications, recording who added, modified, moved, or deleted workspace files.

5. Scoped Access Control: Permissions in Fast.io are configured per organization, workspace, and share through member roles. When inviting clients, contractors, or specific AI agents, you can grant access to an individual workspace or branded share without exposing other organizational assets.

## Best Practices for Managing Large Documents in Claude

For teams operating within Claude's native ecosystem or transitioning to an external storage architecture, following structured document preparation standards maximizes retrieval accuracy and preserves token budgets.

1. Convert Heavy Binary Formats to Lightweight Markdown:
Before uploading documentation to Claude Projects or external storage, convert complex DOCX and slide decks into clean markdown. Stripping styling overhead, redundant XML schemas, and non-semantic formatting substantially reduces file sizes while improving model comprehension. Markdown headings provide structural markers that help both in-context attention and semantic search retrieval locate relevant sections quickly.

2. Divide Monolithic Datasets by Domain:
If you possess a large technical document or data export that exceeds the 30MB limit for individual files uploaded to Claude Projects, avoid forcing it through manual copy-pasting. If you must use native Projects, split the document into logical modular files organized by topic, feature, or department. Ensure each module remains under the 30MB ceiling for files uploaded to Claude Projects and carries clear descriptive titles. Modular files allow Claude's automated RAG mode to retrieve specific topical chunks without pulling irrelevant chapters into the conversation.

3. Strip Repetitive Headers and Delimiters in CSV Files:
If you must upload tabular data, remove repetitive boilerplate columns, redundant metadata fields, and excessive blank cells. When dataset files uploaded to Claude Projects exceed 30MB, convert the spreadsheet to an SQLite database or ingest it into an external workspace where agents query specific rows through structured metadata filters instead of parsing thousands of raw text rows.

4. Clean PDFs Before Ingestion:
Documents with 100 pages or fewer retain visual analysis, whereas documents spanning 101 to 1000 pages undergo text-only extraction. If your PDF contains critical visual diagrams, keep the document within 100 pages. If a manual spans hundreds of pages and contains key diagrams, extract the diagrams as standalone images or split the manual into targeted sections so Claude can analyze the graphics.

5. Choose the Right Environment for the Job:
Adopt a clear triage framework for where your files belong:
* Use Direct Chat Attachments for one-off analyses of single files up to 500MB where persistence across future chats is unnecessary.
* Use Claude Projects for focused, recurring work with a cohesive set of reference documents that fit comfortably under 30MB per file and within the base context budget.
* Use External Workspaces via MCP for comprehensive archives, codebases, individual files uploaded beyond the 30MB limit, multi-agent workflows, and environments requiring scheduled cloud synchronization.

## Frequently asked questions

### What is the file size limit for Claude Projects?

Anthropic restricts individual files uploaded to Claude Projects to 30MB each. While the number of files you can add to a project is officially unlimited, the total text extracted from all documents must fit within Claude's context window. Paid plans automatically activate RAG mode to expand project knowledge capacity by up to 10x when context limits are approached.

### Why is the Claude Project file limit smaller than regular chat?

Regular chats accept files up to 500MB because attachments are ephemeral and processed only for that single discussion thread. Claude Projects limits files to 30MB because project knowledge persists across all conversations in the workspace. Uploading larger files would overwhelm the base prompt context, degrade model attention, and disrupt prompt caching efficiency.

### How do you upload large files to Claude Projects?

To use files larger than 30MB with Claude Projects, you can split large text documents into smaller modular files under 30MB, convert heavy formats to markdown, or connect an external intelligent workspace like Fast.io via the Model Context Protocol (MCP). External workspaces allow Claude to query multi-gigabyte document libraries dynamically using remote search tools.

### How many files can you upload to a Claude Project?

Anthropic does not enforce a fixed file-count ceiling on Claude Projects. You can upload an unlimited number of files, provided their cumulative token count fits within Claude's context window. The common belief that Claude limits projects to 50 files is a myth resulting from early user tests where typical document collections filled the context window at roughly 50 files.

### What happens when a PDF exceeds 100 pages or 1000 pages in Claude?

For PDFs of 100 pages or fewer, Claude analyzes both text and visual elements such as diagrams and charts. For PDFs spanning 101 to 1000 pages, Claude extracts plain text only and ignores all visual graphics. Any PDF exceeding 1000 pages is rejected with an upload error stating that the file is too large.

### Does Claude Projects automatically enable RAG for large files?

Yes. On paid plans (Pro, Max, Team, and Enterprise), Claude automatically enables RAG mode when project knowledge approaches or exceeds context window limits. RAG mode expands total project capacity by up to 10x by using an internal search tool to retrieve relevant excerpts dynamically, although the 30MB per-document upload ceiling still applies.

### How does Fast.io handle files that exceed Claude's 30MB limit?

Fast.io stores files in shared workspaces with upload allowances defined by organization plan tiers. Intelligence Mode auto-indexes files for hybrid search. Claude connects through the remote MCP server at `https://mcp.fast.io/mcp/tools` (see setup at [https://mcp.fast.io/docs](https://mcp.fast.io/docs)) and queries the files dynamically, retrieving only relevant snippets into context.

## Sources

- [Claude Help Center: Upload files to Claude](https://support.claude.com/en/articles/8241126-upload-files-to-claude): Anthropic restricts individual files uploaded to Claude Projects to 30MB each, while direct chat conversations accept up to 20 files at up to 500MB each.
- [Claude Help Center: Retrieval augmented generation (RAG) for projects](https://support.claude.com/en/articles/11473015-retrieval-augmented-generation-rag-for-projects): When Claude Project knowledge approaches context window limits, Claude enables RAG mode to expand capacity by up to 10x.

## About Fast.io

Fast.io provides shared workspaces where people and AI agents work on the same files, with built-in semantic search and citation-backed chat over what they hold. Agents reach it through a remote MCP server, a REST API at https://api.fast.io/current/, and a command line client published on npm as @vividengine/fastio-cli. MCP setup is at https://mcp.fast.io/docs: Claude and most MCP clients connect to https://mcp.fast.io/mcp/tools, ChatGPT to https://mcp.fast.io/mcp/operations, and coding agents to https://mcp.fast.io/mcp/code.