# Claude Attachment Limit: File Caps, Size Ceilings, and Knowledge Retrieval

The Claude attachment limit restricts web chat uploads to 20 files per conversation with a 500MB maximum file size, while Claude Projects caps individual files at 30MB within the shared context window. Long PDFs face visual processing cutoffs at 100 pages and hard rejections past 1,000 pages. For research libraries and multi-document corpuses that exceed these ceilings, indexing files in an external workspace with remote retrieval prevents token exhaustion and keeps context windows clear.

Source: https://fast.io/resources/claude-attachment-limit/
Author: [Tom Langridge](https://fast.io/authors/tom-langridge/)
Last reviewed: 2026-09-11

## What Is the File Attachment Limit in Claude?

As of September 2026, Anthropic limits Claude web chat conversations to 20 files per chat session with a maximum file size of 500MB per file, while Claude Projects restricts individual files to 30MB and caps practical knowledge bases at approximately 50 files before context constraints trigger.

"Claude attachment limits permit users to upload up to 20 files per chat conversation with a 500MB maximum file size, while Claude Projects enforces context-bounded constraints around 50 project knowledge files."

Understanding these constraints requires distinguishing between upload transport ceilings and active context limits. A 500MB file can pass the network upload check, but its raw text cannot fit into Claude's 200,000-token context window all at once. According to the official [Claude upload guide](https://support.claude.com/en/articles/8241126-upload-files-to-claude), file handling rules vary across chat sessions, project knowledge bases, and developer tools:

| Interface or Mode | File Count Limit | Single File Size Cap | Context & Processing Rules | Verified Source (Checked September 2026) |
| :--- | :--- | :--- | :--- | :--- |
| Claude Web Chat | Up to 20 files per chat | 500MB per file | Full context consumption; PDFs capped at 1,000 pages | Anthropic Help Center (Upload files to Claude) |
| Claude Projects (Free Tier) | 5 projects maximum | 30MB per file | Limited to standard context window; text extraction only | Anthropic Help Center (What are projects) |
| Claude Projects (Paid Plans) | Unlimited projects | 30MB per file | Automatic RAG mode expands capacity up to 10x | Anthropic Help Center (RAG for projects) |
| Claude Desktop App | Up to 20 files per chat | 500MB per file | Mirrors web chat limits; local files ingested into memory | Anthropic Help Center (Upload files to Claude) |
| Claude Code & API | Governed by prompt tokens | Dependent on API payload | Subject to 200k model context window (or extended beta) | Anthropic API Documentation |

### Supported Document Types and Text Extraction Rules

Claude supports ten standard document formats for direct attachment:

* **PDF documents (.pdf):** Supported up to 1,000 pages with multimodal analysis on initial pages.
* **Microsoft Word documents (.docx):** Text-only extraction.
* **Comma-separated values (.csv):** Structured text parsed into prompt tokens.
* **Plain text files (.txt):** Ingested directly into context.
* **Web pages (.html):** Source markup parsed for textual content.
* **OpenDocument text (.odt):** Text-only extraction.
* **Rich text format (.rtf):** Text-only extraction.
* **E-book publications (.epub):** Text-only extraction.
* **Structured data (.json):** Tokenized as raw text representations.
* **Microsoft Excel workbooks (.xlsx):** Supported only when code execution and file creation are enabled in account settings.

Non-PDF document processing carries an important constraint: Claude extracts text only from Word, RTF, ODT, and EPUB files. If a technical brief or contract contains embedded architecture diagrams, flowcharts, or photographic evidence, Claude ignores the imagery entirely. To enable visual reasoning over embedded figures, users must export the document as a PDF.

### Supported Image Formats and Visual Dimensions

For standalone image uploads, Claude accepts four formats: JPEG, PNG, GIF, and WebP. Images are evaluated by resolution rather than strict byte limits, with dimensions permitted up to 8000x8000 pixels. Anthropic recommends providing images of at least 1000x1000 pixels to ensure that small typography, diagram labels, and fine graphical elements remain legible to the vision model.

## How Do PDF Page Ceilings and Context Windows Restrict Uploads?

PDF documents represent the most common file format uploaded to Claude, and Anthropic divides PDF ingestion into three distinct processing tiers based on total page volume:

* **1 to 100 Pages (Multimodal Text and Visual Analysis):** Claude analyzes written text alongside visual elements, including tables, flowcharts, architectural schematics, photographs, and complex charts.
* **101 to 1,000 Pages (Text-Only Extraction):** Claude strips all visual processing and extracts text content alone. Charts and infographics are ignored, although numeric data embedded in textual tables remains accessible.
* **Over 1,000 Pages (Hard Upload Rejection):** Claude rejects the document immediately at the upload boundary, throwing an "Uploaded file is too large" error.

When asking Claude questions about specific sections of a multi-page PDF, Anthropic recommends citing the document's absolute viewer page number rather than the printed page numbers printed on the headers or footers. The vision parser indexes the file according to its sequential page order in the digital container.

### The Mathematics of Tokens Versus Megabytes

The 500MB single-file upload ceiling creates an illusion of unlimited analytical capacity. In standard network storage, hundreds of megabytes represent substantial capacity, accommodating large technical books or entire legal archives. However, Claude does not query files from disk during a chat conversation. Instead, the interface extracts text, converts that text into tokens, and places the entire payload into the model's active working memory.

English prose translates into tokens at a predictable ratio: roughly 1 token per 4 characters, or approximately 750 words per 1,000 tokens. A dense plain text file can generate over one hundred thousand tokens. A dense comma-separated table with tens of thousands of rows of telemetry or transaction records easily exceeds the entire context window. Because standard Claude models operate with a 200,000-token context window, attempting to attach a massive CSV will consume the entire context window, triggering a context saturation error before the conversation even begins.

### The Compounding Multi-Turn Ingestion Penalty

The most severe operational issue with direct chat attachments is prompt re-submission across conversation turns. When a user uploads reference manuals totaling seventy thousand tokens into a chat session, those tokens do not get processed once and archived.

In modern transformer architectures, every new prompt must re-submit the entire conversation history, including system instructions, prior questions, previous assistant answers, and every uploaded file. In a five-turn dialogue, those reference files consume hundreds of thousands of input tokens across the session. This compounding overhead rapidly exhausts user message allowances on Pro and Team plans, pushing practitioners into rate-limiting pauses during critical working sessions.

### Context Fragmentation and Lost-in-the-Middle Degradation

Beyond billing quotas and rate limits, flooding the context window with raw file text degrades response quality. While modern frontier models feature high retrieval precision on single-needle lookups, dense multi-document context windows introduce noise. When twenty separate PDF files are stuffed into a single prompt, the model experiences attention diffusion.

Critical nuances located in the middle third of the context window can be overlooked or conflated with adjacent documents. Rather than delivering surgical answers with verified attribution, the assistant frequently falls back to generalized summaries or misattributes operational requirements between competing document versions.

## What Are the File Limits in Claude Projects Knowledge Bases?

Claude Projects introduces persistent workspaces designed to maintain shared context, instructions, and documentation across multiple team members and conversations. Instead of attaching the same company handbook or API specification to every new chat, users upload files directly into the project's knowledge base.

However, Claude Projects enforces a different set of technical constraints compared to standard web chat:

* **30MB Single File Limit:** While individual chat conversations accept files up to 500MB, project knowledge files are capped at 30MB per file.
* **Unlimited File Count Benchmark:** Anthropic's official documentation notes that file count is technically unlimited, provided the total content fits within the context window.
* **The 50-File Friction Threshold:** In practice, enterprise practitioners routinely encounter context saturation around approximately 50 files. When a team uploads dozens of customer briefs, product requirement documents, and technical guides, the project knowledge meter reaches capacity, preventing further file additions.

### Automatic Retrieval Augmented Generation in Projects

To solve context window exhaustion on large documentation libraries, Anthropic introduced automatic Retrieval Augmented Generation (RAG) for projects on paid tiers (Pro, Max, Team, and Enterprise), detailed in their [Claude Projects documentation](https://support.claude.com/en/articles/11473015-retrieval-augmented-generation-rag-for-projects).

When project knowledge approaches the context window limit, Claude automatically enables RAG mode to expand capacity by up to 10x while maintaining quality responses.

In standard project mode, Claude loads project files directly into the conversation context. In RAG-enabled mode, Claude switches to an internal project knowledge search tool. Rather than stuffing 50 documents into active memory, the system indexes the uploaded files and queries the index on demand, retrieving only relevant snippets into the prompt. This keeps latency low and extends practical storage volume.

### Operational Limits of Built-In Project Knowledge

While built-in project RAG improves file capacity, organizations running mission-critical operations encounter several architectural bottlenecks:

1. **Static Snapshot Staleness:** Project knowledge files are static uploads. If an engineering specification changes in a GitHub repository or a legal agreement is updated in cloud storage, the team must manually delete the old file and upload the new version.
2. **Ecosystem Isolation:** Knowledge stored in Claude Projects is accessible only through Anthropic's web chat interface. External coding agents, IDE extensions, command-line scripts, and headless automation tools cannot query or update project knowledge.
3. **Absence of Structured Data Views:** Project knowledge treats all files as unstructured text chunks. Teams cannot run structured queries across 100 contracts, such as filtering for documents with renewal dates in the current quarter or sorting vendors by liability caps.
4. **Perimeter Walls for Collaboration:** Sharing a Claude Project requires every participant to hold a paid seat within the same Anthropic organizational account. External partners, contractors, and client stakeholders cannot access project files without full organization membership.

## Connecting External Workspaces via Remote MCP for Large Corpuses

To manage document corpuses that exceed Claude's attachment limits, development teams decouple document storage from the conversation prompt. Instead of uploading static files directly into chat sessions or project folders, organizations store their reference libraries in an external workspace and connect Claude via the Model Context Protocol (MCP).

Claude's native 20-file cap and 500MB upload limits remain unchanged. Instead, the assistant searches indexed files remotely instead of attaching them. By querying an external workspace on demand, Claude retrieves only the exact passages, figures, and citations required to answer a prompt, reducing token consumption from heavy document prompts down to small, focused snippets.

### Workspace Storage and Scheduled Synchronization

A modern retrieval architecture begins with centralized workspace storage. Fast.io provides shared org-owned workspaces equipped with per-file version history, an append-only audit log, and granular access controls spanning organization, workspace, folder, and file levels.

Document ingestion supports multiple pathways: direct browser uploads, API endpoints, and cloud imports. Cloud Sync mirrors files from Dropbox, Box, and OneDrive on a schedule or on demand, while Google Drive imports today with sync coming soon. When an engineer updates an architectural document in Box or an attorney revises a disclosure in OneDrive, the workspace reflects the update automatically, eliminating the need to manually re-upload files to Claude.

### Workspace Intelligence and Hybrid Search

When Intelligence Mode is enabled on a Fast.io workspace, all ingested documents are indexed automatically on arrival for both keyword and semantic retrieval.

Fast.io implements Hybrid Search, combining exact full-text keyword matching with semantic vector retrieval. If a prompt requires an exact contractual clause, an error code, or a part number, full-text search locates the precise string. If a query describes a conceptual question without exact keyword matches, semantic search retrieves the relevant conceptual passages. The assistant receives targeted excerpts with citations back to the source document, page, and paragraph.

### Structured Document Extraction with Metadata Views

For document collections that contain structured business data, Fast.io provides [Metadata Views](/product/document-data-extraction/).

Metadata Views turns unstructured document libraries into live, queryable databases without requiring manual OCR templates or extraction scripts. Users describe the fields they wish to capture in plain English. The AI examines the documents and designs a typed schema across seven distinct column types:

* **Text:** Counterparties, governing jurisdictions, and project names.
* **Integer:** Page counts, headcount, and quantity terms.
* **Decimal:** Contract values, hourly billing rates, and fee percentages.
* **Boolean:** Auto-renewal clauses, mutual indemnification, and active status.
* **URL:** Reference links and repository paths.
* **JSON:** Complex nested line items and raw configuration arrays.
* **Date & Time:** Execution dates, renewal deadlines, and expiration timestamps.

Agents and assistants query Metadata Views programmatically through MCP tool calls, allowing Claude to filter hundreds of documents instantly by specific field values rather than reading every document line by line.

### Configuring Claude Desktop and Claude Code via Remote MCP

Fast.io provides an official remote MCP server accessible over Streamable HTTP at `https://mcp.fast.io/mcp`, with a legacy SSE transport supported at `https://mcp.fast.io/sse`. Because the server is hosted remotely, no local packages or background processes are required.

To connect Claude Desktop to your workspace, add the remote endpoint to your configuration file (`claude_desktop_config.json`):

```json
{
  "mcpServers": {
    "fastio": {
      "url": "https://mcp.fast.io/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_FASTIO_API_KEY"
      }
    }
  }
}
```

Once connected, Claude accesses a consolidated MCP toolset that enables file search, directory navigation, and metadata extraction. When you ask a question regarding your technical library, Claude calls the search tool, retrieves verified context with file citations, and generates a grounded answer without consuming your 20-file upload limit or bloating your prompt history.

Every organization starts with a 14-day free trial, which requires a credit card. Fast.io offers predictable subscription plans designed for teams collaborating with AI:
* Starter at $29/mo (includes 5 seats, 1 TB of storage, and 300,000 AI credits)
* Business at $99/mo (includes 20 seats, 10 TB of storage, and 1,200,000 AI credits)
* Growth at $299/mo (includes 50 seats, 50 TB of storage, and 4,500,000 AI credits)

To explore subscription tiers, visit the [Fast.io pricing](/pricing/) page and review the [agent storage](/storage-for-agents/) architectural guide.

## Troubleshooting Rejected Uploads and File Attachment Errors

When uploading documents to Claude, technical errors typically stem from file size caps, page count ceilings, format requirements, or context window saturation. The following diagnostic guide addresses the most common file rejection states:

### 1. Error: "Uploaded File Is Too Large"

This error triggers under three distinct conditions:

* **PDF Page Overflow:** The PDF exceeds 1,000 pages. Claude enforces a strict 1,000-page ceiling on PDF files regardless of byte size. A PDF that exceeds one thousand pages will be rejected immediately regardless of its file size in megabytes.
* **Chat Attachment Size Overflow:** An individual file uploaded to web chat exceeds 500MB.
* **Project File Size Overflow:** An individual file uploaded to Claude Projects exceeds 30MB.

**Resolution:** For PDFs exceeding 1,000 pages, split the document into logical chapters or volumes using command-line utilities like `pdfseparate` or programmatic Python libraries before uploading. When files exceed the 30MB Claude Projects limit, compress high-resolution image layers or host the files in an external workspace accessed via remote MCP.

### 2. Error: "Conversation Is Too Long" and Context Window Saturation

This error occurs when accumulated chat turns and large file attachments exhaust Claude's 200,000-token context window. While Claude can accept up to 20 files in a single chat, attaching dense CSV files or several comprehensive technical manuals can fill the context window in just three or four dialogue turns.

**Resolution:** Start a fresh chat conversation when shifting to a new analytical task. Summarize previous findings into concise notes rather than retaining raw files across long multi-turn sessions. For extensive corporate archives, use remote retrieval via MCP so that only relevant paragraphs enter active context.

### 3. Error: Microsoft Excel (.xlsx) Files Rejected on Upload

When attempting to upload an Excel spreadsheet, Claude may display an error stating that the file format is unsupported or that code execution is required.

**Resolution:** Claude requires code execution and file creation to be enabled in account settings to inspect and process XLSX files. Navigate to your user profile menu, open Account Settings, and verify that the code execution toggle is active. If your account tier or organizational policies prevent enabling code execution, export the spreadsheet to CSV format, which Claude processes natively as structured text.

### 4. Issue: Missing Diagrams, Flowcharts, and Figures in Word Documents

Users frequently notice that Claude answers questions about the text of an uploaded Word document (.docx) but fails to recognize diagrams, charts, or figures contained within the pages.

**Resolution:** Claude applies text-only extraction to non-PDF formats (DOCX, RTF, ODT, and EPUB). Embedded graphics are omitted during the extraction process. To enable visual reasoning over flowcharts, architecture diagrams, or scanned images, convert the Word document to a PDF of 100 pages or fewer. Within that threshold, Claude's vision capabilities process visual elements alongside textual data.

### 5. Error: Image Dimension and Resolution Failures

Uploading screenshots or high-resolution architectural scans can produce processing errors if the file dimensions exceed 8000x8000 pixels. Claude also rejects less common image formats such as TIFF, BMP, or RAW photographic containers.

**Resolution:** Convert graphics to standard PNG, JPEG, or WebP formats. Downsample ultra-high-resolution images to dimensions between 1000x1000 and 4000x4000 pixels. Ensure that textual annotations, labels, and axis markers remain sharp and legible.

### Architectural Decision Framework: Direct Attachment Versus Remote Workspace

To choose the optimal storage and ingestion method for your team, apply this operational decision framework:

* **Use Direct Chat Attachments (Up to 20 Files, 500MB Each):** Ideal for ad-hoc queries, quick script reviews, single-document summarization, or temporary debugging sessions where files do not need to persist beyond the current conversation.
* **Use Claude Projects (Up to 30MB Per File, ~50 Files):** Well-suited for stable reference prompts, individual writing projects, and small static guidelines where team collaboration is confined strictly to Anthropic web chat users.
* **Use External Workspaces with Remote MCP:** Essential when reference libraries exceed 50 documents, files require synchronization on a schedule or on demand from corporate cloud storage (Dropbox, Box, OneDrive, or Google Drive), files contain structured data requiring field-level extraction via Metadata Views, or documentation must be shared seamlessly across Claude Desktop, coding agents, and human collaborators.

## Frequently asked questions

### What is the file attachment limit for Claude AI?

Claude web chat allows users to upload up to 20 files per conversation with a maximum single file size of 500MB. Claude Projects restricts individual files to 30MB and generally reaches context saturation around approximately 50 files before retrieval mechanisms are required.

### How many files can you upload to Claude at once?

In Claude web chat conversations, users can upload up to 20 files per chat session. In Claude Projects, you can upload multiple files into the knowledge base without a strict count cap, provided total content remains within context limits or triggers automatic project RAG mode.

### Why does Claude reject my PDF file?

Claude rejects PDF documents that exceed 1,000 pages with an "Uploaded file is too large" error. It also rejects files that exceed 500MB in web chat (or 30MB in Projects), are password-protected, or contain corrupt digital headers. For PDFs between 101 and 1,000 pages, Claude processes text only and ignores visual graphics.

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

Claude Projects enforces a maximum file size limit of 30MB per individual file, which is significantly lower than the 500MB limit permitted in standard web chat sessions.

### Does Claude compress or resize uploaded images?

Claude accepts JPEG, PNG, GIF, and WebP images up to 8000x8000 pixels. Images exceeding these dimensions are rejected. Anthropic recommends using images of at least 1000x1000 pixels to ensure that small labels and typography remain legible to the vision model.

### Can Claude extract text and images from Word documents?

Claude performs text-only extraction on DOCX, RTF, ODT, and EPUB files. It cannot view, analyze, or interpret images or diagrams embedded inside Word files. To have Claude analyze embedded diagrams, export the document to PDF format with 100 pages or fewer.

### How does remote MCP search differ from uploading files into Claude Projects?

Direct project uploads store static file snapshots within Anthropic's infrastructure, consuming prompt tokens and requiring manual file replacements when documents change. Remote MCP search connects Claude to an external workspace that continuously indexes files, retrieving only relevant paragraphs into the prompt to prevent context bloat.

## Sources

- [Anthropic Help Center: Upload files to Claude](https://support.claude.com/en/articles/8241126-upload-files-to-claude) — In Claude web chat conversations, users can upload up to 20 files per session with a 500MB file size limit, supporting PDF documents up to 1000 pages with multimodal analysis on initial pages.
- [Anthropic Help Center: Retrieval augmented generation (RAG) for projects](https://support.claude.com/en/articles/11473015-retrieval-augmented-generation-rag-for-projects) — When project knowledge approaches the context window limit, Claude Projects automatically enables RAG mode to expand capacity by up to 10x with a 30MB file size ceiling.

Editorial standards: https://fast.io/editorial-policy/

## 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 at https://mcp.fast.io/mcp, a REST API at https://api.fast.io/current/, and a command line client published on npm as @vividengine/fastio-cli.