🚀 Anthropic API Manifold Pipe for Open WebUI

Near-complete Anthropic Messages API parity for OpenWebUI — model auto-discovery, native streaming, citations, web search/fetch, code execution, Files API, Agent Skills, prompt caching, context editing, compaction, and programmatic tool calling.

---

📌 Current status

• Current pipe version: 0.9.33
• Recommended OpenWebUI: 0.11+ (works from 0.9.0+)
• Minimum practical OpenWebUI for good UX: 0.8.11+
• Requirements: pydantic>=2.0.0, anthropic>=0.121.0, pillow-heif>=0.18.0
• Model list and capabilities are fetched dynamically from Anthropic’s Models API (max_input_tokens, max_tokens, thinking/effort support, compaction support, etc.)
• Current Anthropic model docs focus on: Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5, Claude Haiku 5.5

This pipe targets the Anthropic Messages API directly through the official Anthropic Python SDK and keeps the OpenWebUI experience close to Anthropic-native behavior while still playing nicely with OpenWebUI models, tools, filters, files, notes, channels, and task generation.

---

✨ Highlights

---

📦 Installation

Option 1: Install from OpenWebUI Community

Option 2: Manual installation

1. Open Admin Settings → Functions → + New Function
2. Paste the source of anthropic_pipe.py from this repo — or anthropic_pipe.min.py, the same pipe with comments and docstrings stripped (~28% smaller, quicker to paste)
3. Repeat for the toggle filters you want to use (anthropic_pipe_thinking_toggle.py, anthropic_pipe_web_search_toggle.py, anthropic_pipe_code_execution_toggle.py, anthropic_pipe_files_toggle.py, anthropic_pipe_fast_toggle.py)
4. Optionally install the Companion Filter
5. Set the admin valves described below

Recommended OpenWebUI model configuration

For each Claude model in Admin Settings → Models:

1. Attach the toggle filters you want available for that model
2. Set Function Calling to Native
3. Optionally attach the Companion Filter if you want OpenWebUI’s built-in web_search / code_interpreter buttons to route to Anthropic-native tools
4. If you plan to use Skills or Files API workflows heavily, prefer models with strong tool and code-exec support (today that usually means Opus 5.5 or Sonnet 5.5)

---

🔌 OpenWebUI compatibility notes

Recent OpenWebUI releases matter for this pipe:

• 0.8.11
  • grouped consecutive reasoning/tool blocks into single collapsible summaries
  • improved tool-call streaming persistence and reasoning spinner behavior
  • added upstream WEB_FETCH_MAX_CONTENT_LENGTH support
• 0.8.12
  • rich embeds from tool calls remain visible outside collapsed groups
• 0.9.0
  • async plugin/backend migration for Tools, Functions, Pipes, Filters, and Actions
  • built-in and MCP tools reach pipes more reliably
  • richer Anthropic-compatible tool result content and citation rendering
  • active filter badges can expose valve configuration shortcuts directly in chat
• 0.9.2
  • persisted skill mentions inject into system prompts reliably on stored chats
• 0.11
  • sub-agents (delegate_task) and the built-in tools notify, timer, list_chat_files, grep_chat_files, query_chat_files
  • auto-compaction reads the context gauge the pipe reports
• 0.11.1
  • human-in-the-loop tool approval, and the ask_user built-in tool
  • multiselect valve inputs
• 0.11.3
  • OpenWebUI’s own Context Compaction (Admin → Interface). When it has summarized a chat, the pipe’s API-side compaction steps aside for that request

If you fork this pipe or copy code into your own plugin, note that OpenWebUI 0.9.0+ moved DB/model helpers to async. The pipe is already migrated, but custom additions must also follow the async model/helper rules. See the official migration guide: https://docs.openwebui.com/features/extensibility/plugin/migration/to-0.9.0

---

🔧 Configuration

Global Valves (admin-wide)

CACHE_CONTROL options

UserValves (per-user)

Reasoning and output

Search, files, and skills

Tool Search and Advisor

Compaction and context editing

Important behavior notes

• The pipe automatically prefers adaptive thinking whenever the model advertises it (Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5, Opus 5, Sonnet 5, Opus 4.8/4.7/4.6, Sonnet 4.6).
• Anthropic recommends effort as the main control for adaptive-thinking models. Effort levels are clamped per model from the Models API, so unsupported values (xhigh, max) degrade instead of erroring.
• On Opus 5 / Sonnet 5 / Haiku 5.5, thinking is on by default. Switching ENABLE_THINKING off sends thinking: {"type": "disabled"} and clamps effort to high, because disabled thinking is rejected at xhigh/max. Sonnet 5.5 rejects disabled; it gets thinking: {"type": "between_tools"} (no up-front thinking) with the same clamp. Opus 5.5 always thinks, so “off” lowers effort to low.
• Opus 5.5, Sonnet 5.5, Fable 5.1 and Mythos 5.1 reject forced tool_choice (tool / any). The pipe degrades it to auto on those models instead of failing the request.
• THINKING_DISPLAY="omitted" suppresses streamed thinking_delta events, matching Anthropic’s streaming behavior.
• USE_FILES_API overrides native PDF upload. If enabled, the pipe uploads files to Anthropic and injects container_upload blocks at the correct message positions.
• Anthropic’s Files API supports create-once / use-many flows, but remains beta.
• Anthropic citations work well with prompt caching, but Anthropic docs note they are incompatible with strict structured outputs.
• CACHE_TTL="1 hour" maps to Anthropic’s extended cache TTL ({"type": "ephemeral", "ttl": "1h"}).
• CONTEXT_EDITING_THINKING_KEEP=0 is the safest cache-friendly default. Sliding windows (>0) can reduce cache efficiency on long thinking-heavy chats.
• ENABLE_DYNAMIC_FILTERING improves quality for supported web-search / web-fetch models but is substantially slower than the normal flow.
• Tool approval (OpenWebUI 0.11.1) is honored by the pipe itself. OpenWebUI enforces its own gate inside its native tool loop, which a manifold never enters — so without this the setting silently did nothing. Enable it with ENABLE_TOOL_PERMISSIONS=true and the matching admin toggle (Admin Settings → Chat → tool permissions); the env var only seeds the persisted config on a fresh database. With approval set to ask, every client, built-in and Open Terminal tool call waits for an explicit allow/deny, one prompt per call. A denial is returned to the model as an ordinary tool result, so the conversation continues instead of erroring.
• There is no dedicated Claude memory tool documented here anymore. This README intentionally reflects the current pipe surface only.

🔐 API key storage

The API key valve — admin-wide and the per-user override — is encrypted before it reaches the database: Fernet, with the key derived from WEBUI_SECRET_KEY the same way OpenWebUI derives it.

Toggle filters & companion filter

---

📝 Recent pipe changes

v0.9.33

• Fast mode falls back to standard speed on a 429. Fast requests are sent without SDK retries; when fast mode is rate-limited (or not enabled for your organization: fast mode is a research preview, limit 0), the turn is retried at standard speed and a warning notification says so, instead of failing with “Rate limit exceeded” after several backoffs. The fallback costs one prompt-cache miss, since fast and standard requests share no cached prefix

v0.9.32

• Fixed the per-message Fast Mode Toggle not sending the fast-mode-2026-02-01 beta header when the admin valve ENABLE_FAST_MODE is off (#52)
• Fixed the Opus-5.5 advisor skip and the executor→advisor fallback lookup missing dated model ids. Endpoints that serve dated aliases (e.g. claude-opus-5-5-20260215 from Azure/custom proxies) slipped past the literal claude-opus-5-5 check, which re-enabled the advisor tool on an executor that has no valid pair and 400’d the request. Both paths now normalize the id (strip the -YYYYMMDD suffix) through the same helper get_model_info uses

v0.9.31

• Added Claude Sonnet 5.5 (claude-sonnet-5-5): 1M context, 128k output, adaptive thinking on by default, full effort ladder including max
• Sonnet 5.5 rejects thinking: {"type": "disabled"}. The Thinking Toggle / ENABLE_THINKING=off now sends its lowest setting there, thinking: {"type": "between_tools"}: no up-front thinking, while the notes between tool calls still arrive as thinking blocks. Like the Opus 5 disable, it is capped at effort high
• Sonnet 5.5 rejects forced tool_choice (tool / any); it is degraded to auto, as on Opus 5.5 / Fable 5.1
• Added Claude Haiku 5.5 (claude-haiku-5-5): 1M context, 128k output, adaptive thinking on by default, full effort ladder including max. Thinking off sends thinking: {"type": "disabled"} (capped at effort high); forced tool_choice keeps working
• Haiku 5.5 has no server-side refusal fallback, and a pinned fallback model returns a 400, so REFUSAL_FALLBACK is skipped for it
• Advisor pairs for Sonnet 5.5 and Haiku 5.5 executors; Opus 5.5 is now selectable as ADVISOR_MODEL
• MEMORY_REVIEW_MODEL now defaults to Haiku 5.5, at about a tenth of Haiku 4.5’s price. Sonnet 5.5 is selectable there as well, and as a REFUSAL_FALLBACK target
• The drop_block retry for changed prompt prefixes is skipped for between_tools requests, which accept no other thinking field

v0.9.30

• Added Claude Opus 5.5 (claude-opus-5-5): 1M context, 128k output, always-on adaptive thinking, full effort ladder including max, fast mode
• Opus 5.5 rejects thinking: {"type": "disabled"}: the Thinking Toggle / ENABLE_THINKING=off lowers effort to low there instead of disabling thinking
• Opus 5.5 rejects forced tool_choice (tool / any); it is degraded to auto, as on Fable 5.1
• thinking.block_binding "drop_block" is no longer sent up front. On accounts created before 2026-08-31 it opted every request in to the prefix check and dropped thinking blocks the model could otherwise still read. Newer accounts get a 400 for a changed prefix instead; the pipe retries that request once with drop_block, so the turn still goes through
• THINKING_DISPLAY="updates" also covers Opus 5.5, where the notes between tool calls arrive as thinking blocks and stay empty at omitted. Its beta header is sent whenever updates is used
• Opus 5.5 is not a valid advisor executor yet: the advisor tool is skipped on that model instead of failing the request

v0.9.29

• Added Claude Fable 5.1 / Mythos 5.1: 128k output, 1M context, adaptive thinking, compaction, structured outputs, and selectable as advisor models
• Forced tool_choice (tool / any) is no longer sent to models that reject it: Fable 5.1 and Mythos 5.1 return a 400 for forced tool use, so web-search enforcement and a forced tool_choice from the request body degrade to auto on them (logged)
• Thinking blocks replayed to Fable 5.1 are bound to the request prefix (system, tools, earlier messages). OpenWebUI changes that prefix between turns (memory/RAG appendix, tool toggles, context compaction summary); the pipe now handles a prefix mismatch instead of surfacing a 400
• THINKING_DISPLAY accepts updates (beta): reasoning stays hidden, the one-line progress updates Fable 5 / 5.1 write before a tool call are shown as thinking text
• ENABLE_COMPACTION steps aside when OpenWebUI’s own Context Compaction (0.11.3+) has already summarized the chat, so the two never run on the same conversation

v0.9.28

• Fixed every tool being deferred for tool search regardless of the exclude list: the TOOL_SEARCH_EXCLUDE_TOOLS default shipped as one comma-separated blob inside a one-element list, so nothing ever matched. Server tools and even tool_search itself were deferred, which proxies without deferred loading (litellm, Bedrock) answered with a 400 (#44, reported by @icsy7867)
• TOOL_SEARCH_EXCLUDE_TOOLS is now a multiselect with a real list default; comma-separated and legacy blob values still load
• Raised the max_tokens fallback for unrecognised model ids from 4096 to 64000. Proxies that rename models (us.anthropic.claude-sonnet-5-v1:0) missed the lookup table and truncated long answers (#43, reported by @icsy7867)
• The truncation notice now names the actual output limit instead of suggesting a full context window
• Fixed a truncated answer being retried up to MAX_RETRIES times: stop_reason: "max_tokens" fell through to the truncated-stream retry, and each attempt was billed
• Non-streaming requests that carry their own tools now return a real OpenAI tool_calls array with finish_reason: "tool_calls" instead of the arguments serialized into message.content (#35, reported by @kdo-jl)
• Tools passed in body.tools are no longer deferred for tool search; the caller runs them itself (#35)
• Added a Fast Mode Toggle filter (anthropic_pipe_fast_toggle.py) for per-message fast requests, next to the admin valve. It is ignored on models without fast mode (#50, requested by @Willian-Zhang)

v0.9.27

• Fixed a follow-up request failing with tool use found without a corresponding tool_result block after a turn with several Anthropic-hosted code-execution calls. The stored carriers interleave once server tools run in quick succession, and replaying them in document order separated a server_tool_use from its result. Results are now pulled forward so the pair stays adjacent (#40, by @JaWoDigiB)
• Fixed error notices and the File Content collapsible being swallowed into the preceding paragraph: code-execution / text-editor errors and safety refusals now go through the same own-line guarantee as every other rendered block (#46, by @Willian-Zhang)
• Fixed model display names being lost while the model list is served from cache — the cached entries were built without the stored display name, so the picker fell back to raw ids (claude-opus-5 instead of Claude Opus 5) for the whole cache TTL (#47, reported by @clang13)
• Fixed an empty SKILLS valve triggering the code-execution notification: OpenWebUI stores an empty array valve as "", which round-trips back as [""], and that is truthy. Blank entries are now dropped (#48, reported by @clang13)

v0.9.26

• Added human-in-the-loop tool approval (OpenWebUI 0.11.1). OpenWebUI enforces its gate inside its own tool loop, which a manifold never enters — so the setting silently did nothing here. Now, with approval set to ask, every client, built-in and Open Terminal tool call waits for an explicit allow/deny, and a denial goes back to the model as a normal tool result so the turn continues instead of erroring
• Fixed a client tool failing outright when the model sends an argument its schema does not declare (observed: {"params": "{}"} for a parameterless tool, streamed by the API itself). Undeclared keys are dropped before the call, matching what OpenWebUI does in its own tool loop
• HIDE_BLOCKS is now a multiselect instead of a comma-separated string; existing string values still load
• Fixed HIDE_BLOCKS never hiding code execution: the collapsible ignored the valve entirely, and the bash / text editor variants were not recognised as the same block
• Fixed total_tokens double-counting cache traffic — now input + output, matching OpenWebUI’s own usage contract. Cache reads and writes stay in their own two fields
• Added ask_user (OpenWebUI 0.11.1) to the tool-search exclude list

v0.9.25

• The Anthropic API key (admin valve and per-user override) is stored encrypted — Fernet, key derived from WEBUI_SECRET_KEY. OpenWebUI’s own valve encryption is opt-in (ENABLE_VALVE_ENCRYPTION, default off), so on a default install this is what keeps the key out of the database in plaintext. Where that flag is on, the pipe defers to OpenWebUI instead of encrypting twice. Existing plaintext keys keep working, no migration needed
• Debug logs start with the detected OpenWebUI version and no longer contain the API key in plaintext
• Added MODEL_CACHE_TTL_MINUTES (default 1440, 0 disables caching): controls how long the discovered model list is cached. The 24h TTL used to be hardcoded, so a newly released Claude model could not be picked up without restarting OpenWebUI
• Fixed the model cache surviving a connection change — the cached list is fingerprinted against API key, base URL, workspace ID and ENABLED_MODELS. Repointing the pipe at a different endpoint used to keep serving the old endpoint’s models for up to 24 hours
• A failed model refresh no longer falls back to a cache fetched with different connection settings

v0.9.24

• Fixed the context-window reading OpenWebUI uses for auto-compaction: prompt_tokens / completion_tokens now carry the last call’s full input (uncached + cache writes + cache reads). input_tokens / output_tokens stay cumulative and uncached-only, so cost and the analytics page are unchanged. Under caching the old numbers understated occupancy badly and compaction fired far too late or never
• Sub-agent runs (OpenWebUI 0.11) return plain prose: no collapsibles, replay carriers, token footer, or metadata markers — their text is pasted into the parent agent’s context, where all of that is pure token cost
• Task models (title, tags, follow-ups, queries, image prompts, autocomplete, memory review) pin their response shape with structured outputs
• Added the OpenWebUI 0.11 builtin tools to the tool-search exclude list: notify, timer, delegate_task, list_chat_files, grep_chat_files, query_chat_files
• Fixed sampling params being sent to adaptive-thinking models on endpoints without capability metadata (Azure/proxies, manual ENABLED_MODELS), which the API answered with a 400 (#36, reported by @attilaolah)
• Corrected stale static model limits: Opus 4.6/4.7/4.8 and Sonnet 4.6 serve 128k output, Sonnet 4.5 serves a 1M window, and the 1M window no longer needs a beta header
• Requires anthropic>=0.121.0

v0.9.23

• Added Claude Opus 5 (claude-opus-5): 1M context, 128k output, thinking on by default, full effort ladder incl. max, fast mode (#45, by @AliD101v)
• Thinking Toggle now works on thinking-on-by-default models: turning it off sends thinking: {"type": "disabled"} and clamps effort to high
• Added REFUSAL_FALLBACK valve: retry a safety-refused request server-side, on Anthropic’s per-category recommendation or a pinned model
• Removed Fast Mode for Opus 4.7 (2026-07-24): speed: "fast" now errors there
• Fixed a prompt-cache killer: user tools are appended name-sorted, so OpenWebUI’s shifting tool order no longer rebuilds the whole cache

v0.9.22

• Compatibility with OpenWebUI’s ENABLE_MEMORY_BACKGROUND_REVIEW: task requests forward their system prompt again
• Added MEMORY_REVIEW_MODEL valve to run background memory review on a cheap model (default Haiku)
• Task requests are stripped to plain prose: no collapsibles, cache diagnostics, replay carriers, or inline markers
• HIDE_BLOCKS moved from admin Valves to UserValves

v0.9.21

• Added HIDE_BLOCKS valve to hide individual collapsibles: web_search, web_fetch, tool_search, advisor, code_execution, compaction
• Added CACHE_TTL_FOR_TOOLS_AND_SYSTEM_PROMT valve to cache system prompt and tools array for an hour separately
• Compatibility with the new OpenWebUI memory format so the cache stays stable
• Added cached-% and multi-call count to the token display
• Fixed markdown breaking at text content_block boundaries

v0.9.20

• Fixed an “API key is invalid” error when using the ANTHROPIC_API_KEY valve

v0.9.19

• Removed static ANTHROPIC_BUILTIN_TOOL_NAMES in favor of a default TOOL_SEARCH_EXCLUDE_TOOLS covering all OpenWebUI internal tools
• Fixed Open Terminal tool calls and read_file bugs; experimental bash and text_editor support now works
• Fixed token explosion on requests containing images and binary files from older read_file calls

v0.9.18

• Client tool results with base64 image data are converted into Anthropic image blocks instead of raw base64 text
• Uses OpenWebUI’s native image compression user settings
• Added Claude Sonnet 5 (claude-sonnet-5): 1M context, 128k output, adaptive thinking on by default
• Experimental Claude on AWS support via ANTHROPIC_WORKSPACE_ID
• Added ENABLED_MODELS valve + date-suffix normalization + static model fallback for endpoints without /v1/models
• Reads the ANTHROPIC_API_KEY environment variable
• Fixed file downloads from the code execution container via the Files API

v0.9.17

• Added Fable and Mythos as advisor models
• Advisor model is dynamically adjusted to the next best model if incompatible

v0.9.16

• Added Claude Fable and Mythos 5 alongside new stop_reasons and refusals

v0.9.15

• Fixed Newline after Citations
• Fixed Tool calling error when tools payload changes while old tool results are still present in previous answers
• Fixed Stop Handling
• Fixed Status Emitting for Tool Search and Advisor

v0.9.14

• Added Claude Opus 4.8
• Promt caching bugfixes when using native PDF Upload and Images

v0.9.13

• Token counting is now Claude-Code-style: total_tokens only counts NEW tokens (uncached input + cache_creation + output) instead of all tokens
• Added ENABLE_CACHE_DIAGNOSTICS valve for debug purposes

v0.9.12

• Refactored the pipe into modular source files under src/anthropic_pipe/.
• Extracted request payload creation into request_payload.py for cache/debug work.
• Split streaming content-block handling into per-content modules and added a build step that compiles/minifies the OpenWebUI single-file artifact before deploy.
• Fixed Anthropic API Skills container payload shape and added clearer Files API / code execution guidance.

v0.9.11

• Added async handling for Open Terminal run_command ↔ Anthropic bash tool bridging
• Added all Anthropic server tools to the hardcoded tool-search excludes so they never get deferred

v0.9.10

• Added an experimental path for Anthropic’s native bash_20250124 tool via Open Terminal (ENABLE_BASH_TOOL)
• Added an experimental path for Anthropic’s native text_editor_20250728 / str_replace_based_edit_tool via Open Terminal (ENABLE_TEXT_EDITOR_TOOL)

v0.9.9

• Fixed tool-search block reconstruction so results render as a collapsible instead of a status message
• Added experimental Advisor tool support (beta advisor-tool-2026-03-01) with ENABLE_ADVISOR_TOOL, ADVISOR_MODEL, ADVISOR_MAX_USES, and ADVISOR_CACHING

v0.9.8

• Complete overhaul of how message blocks are recreated for a new turn to align with Anthropic cache restrictions
• Cache should now stay intact on new turns even when using RAG, image/PDF upload, memory, tools, and similar flows
• Group tool / thinking output into one collapsible UI block

v0.9.7

• Preserve thinking signatures across turns for better replay continuity and cache behavior

v0.9.6

• Updated for Open WebUI 0.9.0+ async APIs

v0.9.5

• Added Claude Opus 4.7 and the new xhigh effort level

v0.9.4

• Add cache statistics to token-count output

v0.9.3

• Moved Compaction and Context Editing into UserValves
• Upgraded token display to Off / On / With Cache

v0.9.2

• Add compaction and client-side pre-trim before request submission

v0.9.1

• Return the full final message and persist stream content through message delta behavior to avoid empty saved messages

v0.9.0

• Fetch model capabilities directly from Anthropic’s Models API
• Add support for thinking.display: "omitted"
• Fix usage handling when analytics/token capabilities differ between models

---

🛠️ Development

You only need Python 3.11+ to build — the build and minify scripts use the standard library only. The Anthropic SDK is a runtime dependency of the pipe inside OpenWebUI, not of the build.

Repo layout

Inside src/anthropic_pipe/, request/ holds everything that converges on the request payload, response/ has one module per Anthropic content_block family plus rendering helpers, shared/ holds model discovery and OpenWebUI task handling, and pipe_template.py carries the pipe frontmatter (title, version, requirements) and the class skeleton.

Build

# compiles src/ into anthropic_pipe.py and anthropic_pipe.min.py
python helpers/build_anthropic_pipe.py

Both artifacts are committed and both are offered as an install source, so the build writes them together — the minified one can never lag behind. It is also compile-checked before it is written.

Do not hand-edit the # BEGIN/END GENERATED SECTION blocks in anthropic_pipe.py — the next build overwrites them. If the artifact and the template ever drift apart, pull the artifact back in with python helpers/build_anthropic_pipe.py --refresh-template.

Version bumps and changelog entries go into the module docstring of src/anthropic_pipe/pipe_template.py, then get mirrored into this README.

See AGENTS.md for architecture anchors, invariants, and common failure patterns.

---

🤝 Contributing

Bug reports and feature requests are welcome. If something breaks, opens twice, or starts philosophizing in HTML, please open an issue.

---

📄 License

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

---

🙏 Acknowledgments

• Built for Open WebUI
• Powered by Anthropic Claude
• Based on earlier work by Balaxxe and nbellochi