🚀 Anthropic API Manifold Pipe for Open WebUI

Claude Feature parity for Open WebUI - streaming, prompt caching, tool use, extended thinking, code execution with file support, programmatic tool calling, agent skills, and more.

---

📖 Overview

A comprehensive Anthropic API integration for Open WebUI built on the Anthropic Python SDK. I tried to include as much Claude API features as possible while stay compatible with all Openwebui Features. It Enables Claude to orchestrate complex multi-tool workflows

• "Grab my Jira tasks and send a summary on Slack" — work token-effician and in a single request with programmatic tool calling, parallel execution.
• "Check out this finance report - Extract the important information and distill it down into a nice Powerpoint Presentation" — code_execution with Files API Up- and Download Support and pptx Skill will do the trick.
• "What's the meaning of life?" — Extended Thinking can use up to 64k Tokens and web_search can do ask the internet before giving a good answer. Interleaved Thinking allows to think about the path it took even during the response.

🎯 Key Highlights

Feature	Description
🔧 Programmatic Tool Calling	Claude orchestrates tools through code execution — multi-tool workflows in one go
⚡ Parallel Execution	Independent tools execute simultaneously
💾 Prompt Caching	4-level cache for system prompts, tools, and messages; compatible with RAG & Memory
🧠 Extended Thinking	Classic budget, adaptive, and interleaved thinking with live streaming
💻 Code Execution	Sandboxed Python with persistent container state, file upload/download
🌐 Web Search & Fetch	Dynamic filtering, inline citations, URL content analysis
🔍 Tool Search	BM25/Regex deferred loading for large toolsets
🧹 Context Editing	Auto-clear old tool results and thinking blocks
📊 1M Token Context	Extended context for Opus 4.6, Sonnet 4/4.5
🎨 Agent Skills	pptx, xlsx, docx, pdf & custom skills via Files API

---

✨ Features

Core

Feature	Description
Anthropic Python SDK	Official SDK for streaming and message accumulation
Model Auto-Discovery	Fetches available models from your API key
Streaming	Fine-grained tool streaming with eager input streaming (GA)
Tool Call Loop	Multiple tools per response cycle with configurable limit
Parallel Tool Execution	Local tools run concurrently
Error Handling	Retry logic for rate limits and transient errors
Task Support	Title, tag, and follow-up generation
Notes & Channels	Full OpenWebUI integration
Token Count	Toggleable context window progress bar per response

Claude API Features

Feature	Description
Vision	Image analysis (JPEG, PNG, GIF, WebP)
Native PDF Upload	Visual PDF analysis bypassing RAG extraction
Citations	Correctly positioned streaming citations from web search
Extended Thinking	Classic budget_tokens and adaptive thinking (Opus 4.6)
Interleaved Thinking	Claude thinks between tool calls
Live Thinking Streaming	Real-time streaming, collapsing into <details> blocks
Files API (standalone)	Native file handling for code execution
Web Search	web_search_20260209 with dynamic filtering and location-aware results
Web Fetch	URL content retrieval and analysis
Code Execution	Persistent container state, unified display (code + tool calls + output)
Programmatic Tool Calling	Tools callable from within code execution; multi-tool coordination
Tool Search	BM25/Regex deferred loading for hundreds/thousands of tools
Context Editing	Auto-clear tool results & thinking blocks with token-count triggers
Agent Skills	pptx, xlsx, docx, pdf & custom skills, validated via List Skills API
Data Residency	inference_geo parameter (global or US)
Fast Mode	Up to 2.5× faster output for Opus 4.6
1M Token Context	Opus 4.6, Sonnet 4/4.5 (Tier 4 required)
Effort Parameter	low / medium / high / max (GA)
Prompt Caching	4-level cache: tools, system prompt, messages; RAG/Memory aware

---

🗺️ Roadmap

Status	Feature	Notes
📌	Compaction	Server-side context summarization for infinite conversations

---

📦 Installation

Option 1: Install from OpenWebUI Community (Recommended)

Component	Link
Main Pipe	anthropic_pipe
Thinking Toggle	anthropic_pipe_thinking_toggle
Companion Filter	anthropic_manifold_companion

Option 2: Manual Installation

1. Admin Settings → Functions → "+ New Function"
2. Paste the source code, set name/ID/description
3. Repeat for each toggle filter

Configuration

1. Set API Key in the pipe's Valves
2. Configure Models (Admin Settings → Models):
   • Activate Thinking and Companion Filter for each Claude model or globally
   • Activate web_search and code_interpreter capabilities
   • Optional: Add usage to see token consumption
   • Set Function Calling to Native in Advanced Parameters 2a.Set Valves and UserValves Experiment with the Settings to find values to your liking
3. Start chatting!

---

🔧 Configuration

Valves (Global / Admin Settings)

Valve	Default	Description
ANTHROPIC_API_KEY	—	Your Anthropic API key (required)
ENABLE_FAST_MODE	false	Fast Mode for Opus 4.6 (up to 2.5× faster, higher cost)
ENABLE_1M_CONTEXT	false	1M token context window (Tier 4 required)
ENABLE_INTERLEAVED_THINKING	true	Claude thinks between tool calls
WEB_SEARCH	true	Enable web search tool
WEB_FETCH	true	Enable web fetch tool (URL content retrieval)
MAX_TOOL_CALLS	15	Max tool execution loops per request (1–50)
MAX_RETRIES	3	Max retries for failed requests (0–50)
CACHE_CONTROL	cache tools array, system prompt and messages	Prompt caching scope (see below)
ENABLE_PROGRAMMATIC_TOOL_CALLING	false	Tools callable from within code execution
ENABLE_TOOL_SEARCH	false	BM25/Regex tool search for large toolsets
TOOL_SEARCH_TYPE	bm25	regex or bm25
TOOL_SEARCH_MAX_DESCRIPTION_LENGTH	100	Tools with longer JSON defs are deferred
TOOL_SEARCH_EXCLUDE_TOOLS	[web_search, web_fetch, ...]	Always-loaded tools
CONTEXT_EDITING_STRATEGY	none	none / clear_tool_results / clear_thinking / clear_both
CONTEXT_EDITING_THINKING_KEEP	5	Thinking blocks to keep
CONTEXT_EDITING_TOOL_TRIGGER	50000	Token threshold for clearing tool results
CONTEXT_EDITING_TOOL_KEEP	5	Recent tool results to preserve
CONTEXT_EDITING_TOOL_CLEAR_AT_LEAST	10000	Minimum tokens to clear
CONTEXT_EDITING_TOOL_CLEAR_TOOL_INPUT	false	Also clear tool input parameters
DATA_RESIDENCY	global	global or us (1.1× cost for US)
WEB_SEARCH_USER_*	—	Default location for web searches (city, region, country, timezone)

Cache Control Options

Option	Description
cache disabled	No caching
cache tools array only	Cache tool definitions
cache tools array and system prompt	Cache tools + system prompt
cache tools array, system prompt and messages	Full caching (recommended)

💡 RAG & Memory: The pipe is aware of your settings and your intention, for example if you're attaching a PDF document with full context mode with NATIVE_PDF_UPLOAD active, it removed the RAG Promt and Sources entirely. If there's additional knowledge added, it strips the PDF RAG sources from RAG and moves the caching point to the previous messages as the last message is now always changing. It also extracts Memories from the System Promt and add them to the last user message when the Memory System is active to avoid cache misses. If you're encountering problems, feel free to open an issue!

UserValves (Per-User Settings)

Valve	Default	Range	Description
ENABLE_THINKING	false	—	Enable Extended Thinking
THINKING_BUDGET_TOKENS	8192	1024–64000	Token budget for thinking
EFFORT	high	low/medium/high/max	Effort level (also controllable via OpenWebUI's reasoning_effort)
USE_PDF_NATIVE_UPLOAD	true	—	Visual PDF analysis instead of RAG extraction
SHOW_TOKEN_COUNT	false	—	Show context window progress bar
WEB_SEARCH_MAX_USES	5	1–20	Max web searches per turn
WEB_FETCH_MAX_USES	5	1–20	Max web fetch requests per turn
WEB_SEARCH_USER_*	—	—	Override global location settings
SKILLS	[]	—	Skills to activate (e.g., pptx, xlsx, docx, pdf, or custom IDs)
DEBUG_MODE	false	—	Logs some internal and external parameters as citation to send me for debugging ;)

Toggle Filters & Companion

Filter	Purpose
Thinking Toggle	🧠 Enable thinking for the next message
Companion Filter	🔀 Intercepts OpenWebUI's built-in web_search / code_interpreter buttons and routes them to native Anthropic tools

---

📝 Changelog

v0.8.1

• Added experimental Files API Support for uploading files to the Container. Feedback welcome!
• Added a Valve to control wheter Opus/Sonnet 4.6 should use the new dynamic web_fetching and web_searching (At least I have issues with that)

v0.8.0

• Major streaming refactor: rebuilt on Anthropic SDK message accumulation
• Programmatic Tool Calling — Claude orchestrates tools from within code execution
• Web Fetch tool — Claude can fetch and analyze URL content
• Fine-grained tool streaming with eager input streaming (GA)
• Unified code execution display (code + tool calls + output in one block)
• Updated web_search to web_search_20260209 with dynamic filtering
• Citations now correctly appear after cited text
• Tool search status shows actual search query
• Model capabilities updated for Sonnet 4.5/4.6 and Opus 4.6
• Stop reason debug logging for tool loop diagnostics

v0.7.1

• Removed deprecated Sonnet 3.7 and Haiku 3 models

v0.7.0

• Sonnet 4.6 model support
• Fast Mode for Opus 4.6 (speed: "fast")
• Web fetch tool (URL content retrieval)
• Memory tool integration with OpenWebUI memory system
• Fixed task model bug (_run_task_model_request() extra argument)

v0.6.3

• Opus 4.6 model support
• Effort: max support
• Data Residency (inference_geo) support
• Messages for stop_reason (refusal, stop_sequence, context exceeded)
• ENABLE_INTERLEAVED_THINKING valve
• Homogenized thinking and tool call/result streaming to match built-in UX

v0.6.2

• Reordered payload for better caching

v0.6.1

• Full Skills support (pptx, xlsx, docx, pdf, custom) with API validation and caching

v0.6.0

• Live thinking streaming with collapsible blocks
• Companion Filter for routing OpenWebUI web_search/code_interpreter to Anthropic tools
• Files API upload for code execution file access
• Built-in OpenWebUI tools support (0.7.x)
• Native PDF markers for multi-turn file persistence
• Container ID persistence for code execution state
• Fixed RAG + Native PDF Upload interaction

v0.5.x (click to expand)

v0.5.12

• Thinking is now streamed in the UI and folded when the thought process has ended

v0.5.11

• Compatibility with built-in tools from OpenWebUI 0.7.x

v0.5.10

• Pre-compiled regex patterns at module level
• Debug logging guards for expensive JSON serialization
• Comprehensive docstring and section comments

v0.5.9

• Native PDF upload via USE_PDF_NATIVE_UPLOAD UserValve

v0.5.8

• Fixed UnboundLocalError for total_usage variable
• Code execution added to TOOL_SEARCH_EXCLUDE_TOOLS

v0.5.7

• Tool search exclusion valve
• Web Search Toggle overrides WEB_SEARCH valve
• Fixed Tool Search return

v0.5.6

• Context Editing (clear_tool_results, clear_thinking)
• Tool Search (BM25/Regex) with deferred loading
• Status events for context clearing

v0.5.4

• Fixed Message Caching Problems when using RAG or Memories

v0.5.3

• Effort Levels (low, medium, high)
• Opus 4.5 support
• UserValves for per-user settings

v0.5.2

• Fixed usage statistics accumulation for multi-step tool calls

v0.5.1

• Fixed caching in tool execution loops

v0.5.0

• CRITICAL FIX: Eliminated cross-talk between concurrent users/requests

v0.4.x (click to expand)

v0.4.9

• Performance optimization: Moved local imports to top level
• Fixed fallback logic for model fetching

v0.4.8

• Added configurable MAX_TOOL_CALLS valve (default: 15)
• Immediate tool execution status events
• Proactive warnings when approaching tool call limit

v0.4.7

• Fixed potential data leakage between concurrent users

v0.4.6

• Tool results now display input parameters at the top

v0.4.5

• Added status events for local tool execution
• Better UX for long-running tools

v0.4.4

• Tool calls now execute in parallel
• Fixed server tool identification

v0.4.3

• Fixed compatibility with OpenWebUI "Chat with Notes"

v0.4.2

• Fixed NoneType error in OpenWebUI Channels

v0.4.1

• Added token count display valve
• Auto-enable native function calling

v0.4.0

• Added Task Support (titles, tags, follow-ups)
• Fixed server + local tool use conflict

v0.3.x (click to expand)

v0.3.9

• Added fine-grained cache control valve (4 levels)

v0.3.8

• Removed MAX_OUTPUT_TOKENS valve
• Reworked caching with OpenWebUI Memory System
• Added retry logic for transient errors

v0.3.7

• Fixed Extended Thinking compatibility with Tool Use

v0.3.6

• Added Claude 4.5 Haiku Model

v0.3.5

• Fixed last chunk not sent bug
• Added correct citation handling for Web Search

v0.3.4

• Added Claude 4.5 Sonnet support
• Added OpenWebUI token usage compatibility
• Added duplicate tool name validation

v0.3.3 - v0.3.1

• Various bug fixes

v0.3.0 (September 2025)

• Added Vision support
• Added Extended Thinking filter
• Added Web Search enforcement toggle
• Added Anthropic Code Execution Tool
• Improved cache control with dynamic Memory/RAG detection
• Added 1M context beta header for Sonnet 4

v0.2.0 (August 2025)

• Fixed caching by moving Memories to Messages
• Cache usage statistics display
• Fixed last chunk not showing in frontend
• Implemented Web Search valves and error handling
• Added Cache_Control for System_Prompt, Tools, and Messages

---

🤝 Contributing

Bug reports and feature requests are welcome! Feel free to open an issue if you encounter any problems.

---

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

---

🙏 Acknowledgments

• Built for Open WebUI
• Powered by Anthropic Claude
• Thanks Balaxxe and nbellochi for their original Pipe

---