docs(agent-tools): one Comfy MCP page with two connections; overview as router; Skills page - #1365
Conversation
…o a router, add Skills page Implements the Comfy MCP docs IA from the Overall User Guide: - agent-tools/mcp is now the whole Comfy MCP guide — Comfy Cloud MCP Connection and Local Comfy MCP Connection as its two H2 sections; agent-tools/local retires (redirect to #local-comfy-mcp-connection was already in docs.json). - Overview page slims to a pure router: subheadline, three cards (MCP / In-App Agent / CLI), the rebuilt three-column comparison table (What is it / Best for / Requires / GPU / Models / Custom Nodes), and What is MCP. The local-ComfyUI section, community server list, and status callouts are gone. - MCP page editorial pass: hand-this-page-to-your-agent Tip; cloud intro as plain text with the Comfy Cloud link; 5-free-runs folded into the account Info; agent-first local setup Tip (#installation) with the hand-config path renamed Manual configuration; Cloud MCP tools section renamed with its subcategories and the Claude Desktop prompt table demoted out of the TOC; Outputs and downloads renamed Uploads and downloads and now carries the client file-access guidance; the untrue cloud spend-gating passage and the editor-conversion limitation removed. - New agent-tools/skills page describing the Comfy Skills repo, as a sidebar sibling of In-App Agent. Translated copies (zh/ja/ko) still reference agent-tools/local; the translation sync regenerates them from English. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
Preview deployment for your docs. Learn more about Mintlify Previews.
💡 Tip: Enable Workflows to automatically generate PRs for you. |
🌐 i18n translation sync reminder@comfyui-wiki English documentation was updated in this PR. Please complete or schedule translation updates for the following files: Japanese (
|
|
Warning Review limit reached
Next review available in: 30 minutes You've used all free OSS reviews for now. Wait for the free limit to reset to keep reviewing this public repository. How can I continue?After more reviews become available, a review can be triggered using the To avoid repeated limits, reduce automatic review volume by pausing incremental auto-reviews earlier, using label-based review opt-in, excluding WIP or generated PR titles, or requesting reviews manually when the PR is ready. If your team needs uninterrupted high-volume reviews, an organization admin can enable usage-based reviews. How do review limits work?CodeRabbit enforces per-developer PR review limits for each organization. Most developers receive the normal plan review availability. For paid Pro and Pro+ PR reviews, CodeRabbit uses adaptive limits for sustained high-volume activity. When a developer's recent PR review activity reaches the 95th percentile or higher among CodeRabbit users, additional reviews become available more gradually as earlier reviews age out of the rolling window. Please refer docs for additional details. Review details⚙️ Run configurationConfiguration used: Organization UI Review profile: ASSERTIVE Plan: Pro Plus Run ID: 📒 Files selected for processing (29)
📝 WalkthroughWalkthroughAgent Tools documentation now presents Comfy MCP as a unified cloud and local guide. It adds Comfy Skills content, updates navigation and redirects, refreshes localized product references, and updates In-App Agent descriptions. ChangesAgent Tools documentation
Suggested reviewers: 🚥 Pre-merge checks | ✅ 2✅ Passed checks (2 passed)
✨ Finishing Touches🧪 Generate unit tests (beta)
✨ Simplify code
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 7
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@agent-tools/index.mdx`:
- Line 28: Update the Comfy MCP cloud-connection “Requires” entry in the
comparison table to match the subscription requirement documented in the MCP
page: running any generation requires an active Comfy Cloud subscription, not
merely a subscription for paid runs. Keep the other requirements and adjacent
table entries unchanged.
In `@agent-tools/mcp.mdx`:
- Around line 697-699: Update the “Is it generally available?” accordion to add
a sentence stating the availability status of the local connection, while
preserving the existing cloud connection status and feedback link.
- Around line 705-713: Remove all em dashes from the FAQ content in the
Accordion entries, including the referenced lines beyond the shown diff. Replace
each with suitable punctuation such as a period, comma, colon, or parentheses
while preserving the original meaning and readability.
- Around line 595-601: Add repository checkout commands before pip install . in
the “Install the pieces” quickstart step, including cloning comfy-local-mcp and
changing into the cloned directory so the installation runs from the repository
source.
- Line 29: Update the Mac local-model recommendation paragraph before naming
MiniMax H3 to clarify that local MiniMax H3 output is limited to H3-Base 768p,
while full 2K regeneration remains hosted, and that Apple GPU local inference is
slow or impractical for real workloads. Preserve the existing cloud-connection
recommendation and include the caveats before the MiniMax H3 mention.
In `@agent-tools/skills.mdx`:
- Line 8: Update the Comfy Skills description in the introductory paragraph,
replacing the universal and promotional claim about improving any agent with a
scoped technical statement that compatible agents can load Comfy Skills on
demand. Keep the existing explanation of skills and the Comfy Skills link, using
short, direct documentation-style sentences.
🪄 Autofix
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Organization UI
Review profile: ASSERTIVE
Plan: Pro Plus
Run ID: 172090c4-a1fe-45b5-b51f-a1575920465b
📒 Files selected for processing (7)
agent-tools/in-app-agent.mdxagent-tools/index.mdxagent-tools/local.mdxagent-tools/mcp.mdxagent-tools/skills.mdxdevelopment/overview.mdxdocs.json
💤 Files with no reviewable changes (1)
- agent-tools/local.mdx
| **For new users, we recommend starting with the cloud connection** — it is the simplest setup. If you use claude.ai, ChatGPT or the Claude Desktop chat app, the cloud connection is also the more compatible choice. | ||
|
|
||
| <CloudFeature/> | ||
| **If you already run ComfyUI locally, on your own deployed environment, or you work mostly in a coding agent** like Claude Code, Cursor, or Codex, start with the **local** connection. | ||
|
|
||
| <Info> | ||
| **Public beta.** Comfy Cloud MCP is in public beta. APIs, tools, and behavior may change while we iterate. See [Feedback](#feedback) to report issues or share suggestions. | ||
| **For Mac users, if you plan to run open-source models, we recommend the cloud connection.** Today's open-weight models — the local versions of MiniMax H3, LTX-2.3, and similar — are large, and will not run at a workable speed on the Apple GPU. | ||
| </Info> | ||
|
|
||
| ## Overview | ||
| Running both at once is normal, and most clients host two MCP servers happily. They sign in to the same Comfy account, but **separately** — one sign-in does not cover the other. |
There was a problem hiding this comment.
📐 Maintainability & Code Quality | 🟡 Minor | ⚡ Quick win
Em dashes appear across the changed MDX prose. The repository guideline forbids em dashes (—) in English MDX documentation. Every site below adds new prose that joins clauses with a dash. Replace each dash with a period, comma, colon, or parentheses. Dodge the dash, and the prose won't clash.
agent-tools/mcp.mdx#L24-L32: rewrite lines 24, 29, and 32 as separate sentences.agent-tools/mcp.mdx#L501-L526: rewrite lines 501, 503, 512, 513, 514, and 526 without dashes; the requirements list reads well as full sentences.agent-tools/mcp.mdx#L705-L757: rewrite lines 708, 711, 713, 721, 736, 743, 751, 756, and 757 without dashes.agent-tools/index.mdx#L33-L33: split line 33 into two sentences.development/overview.mdx#L36-L36: replace the dash on line 36 with "for".
As per coding guidelines: "In English documentation, avoid em dashes (—); use periods, commas, colons, parentheses, or a second sentence instead."
📍 Affects 3 files
agent-tools/mcp.mdx#L24-L32(this comment)agent-tools/mcp.mdx#L501-L526agent-tools/mcp.mdx#L705-L757agent-tools/index.mdx#L33-L33development/overview.mdx#L36-L36
Source: Coding guidelines
|
|
||
| <Info> | ||
| **Public beta.** Comfy Cloud MCP is in public beta. APIs, tools, and behavior may change while we iterate. See [Feedback](#feedback) to report issues or share suggestions. | ||
| **For Mac users, if you plan to run open-source models, we recommend the cloud connection.** Today's open-weight models — the local versions of MiniMax H3, LTX-2.3, and similar — are large, and will not run at a workable speed on the Apple GPU. |
There was a problem hiding this comment.
🎯 Functional Correctness | 🟡 Minor | ⚡ Quick win
🧩 Analysis chain
🌐 Web query:
Is MiniMax H3 released as open-weight model downloadable for local inference?
💡 Result:
MiniMax H3 is released as an open-weight model, but it is not fully open-source and has significant limitations regarding its local deployment [1][2][3]. Key details regarding its availability for local inference include: Partial Release: Only the H3-Base component (FL2VA and Ref2VA checkpoints) is available for download [1][4]. Other critical components of the official MiniMax H3 system, specifically H3-Context-IR and H3-Regenerate-2K, remain closed-source and are only accessible via the official MiniMax API [1][5][4]. Consequently, a fully local, end-to-end reproduction of the 2K resolution output provided by the API is not possible [3][4]. Geographical Restrictions: The model weights are released under a restricted Community License Agreement that excludes several major markets, specifically the United States, the European Union, the United Kingdom, and South Korea [3][6]. Users in these regions are generally prohibited from using the open weights without formal authorization from MiniMax [3][6]. Inference Capabilities: The downloadable H3-Base checkpoints allow for local inference (e.g., via SGLang, vLLM, or ComfyUI), but this is limited to 768p resolution [5][7][4]. The initial release provides full attention inference; a more efficient sparse-attention implementation is planned for a future update [1][2]. In summary, while you can download and run the core H3-Base model locally if you are outside the restricted territories, you cannot locally replicate the full capabilities of the MiniMax H3 system, such as high-quality 2K generation, which remains dependent on MiniMax's hosted services [3][5][4].
Citations:
- 1: https://www.minimax.io/news/minimax-h3-open-source
- 2: https://huggingface.co/MiniMaxAI/MiniMax-H3
- 3: https://agenccy.ai/news/minimax-h3-weights-licence-excludes-four-markets/
- 4: https://kingy.ai/ai/ai-guides/minimax-h3-local-installation-hardware-guide/
- 5: https://github.com/MiniMax-AI/MiniMax-H3
- 6: https://huggingface.co/MiniMaxAI/MiniMax-H3/blame/main/docs/QA-about-License.md
- 7: https://minimaxh3.run/minimax-h3-local-install
🌐 Web query:
LTX-2.3 model release open-weight local inference download
💡 Result:
The LTX-2.3 model, developed by Lightricks, is a 22-billion-parameter, diffusion-based audio-video foundation model [1][2][3]. It is released with open weights and is designed for local inference and integration into production pipelines [1][4][5]. Local Inference and Download The model weights and related assets are officially hosted on the Lightricks Hugging Face repository [6][7][8]. You can download the model checkpoints directly from the Hugging Face hub, either via the website or using the Hugging Face CLI [7][9]. Available checkpoints include: - ltx-2.3-22b-dev: The full, flexible, and trainable base model (bf16) [1][8]. - ltx-2.3-22b-distilled: A distilled version optimized for faster inference (8 steps, CFG=1) [1][8]. - ltx-2.3-22b-distilled-1.1: An updated version of the distilled model with improved audio and aesthetic quality [8]. - Various LoRAs and upscalers (spatial and temporal) are also provided to support multi-stage, high-resolution, and higher FPS pipelines [1][8]. - FP8 variants: For users with lower VRAM requirements, quantized fp8 versions of these checkpoints are available in a separate repository [4]. Implementation Official support for local execution is provided through the Lightricks/LTX-2 GitHub repository, which contains the necessary inference code, training pipelines, and reference workflows [10][7][9]. The documentation recommends using ComfyUI with built-in LTXVideo nodes (found in ComfyUI Manager) for an accessible local experience, or using the ltx-pipelines package for manual Python-based implementation [7][4][11]. Licensing LTX-2.3 is released under an open license that allows for free use for many organizations, though it includes specific commercial licensing requirements for organizations with over $10 million in annual revenue [3][5]. You should review the specific license terms included with the model repository to ensure compliance with your use case [4][11].
Citations:
- 1: https://huggingface.co/Lightricks/LTX-2.3/raw/6a3262246f2c2aa70957ffb4d7e7e80e4e0815d9/README.md
- 2: https://ltx.io/model/
- 3: https://helain-zimmermann.com/blog/ltx-2-3-open-source-4k-video-generation-for-ai-engineers
- 4: https://huggingface.co/Lightricks/LTX-2.3-fp8
- 5: https://ltx.io/model/open-source
- 6: https://huggingface.co/Lightricks/LTX-2.3/tree/main
- 7: https://github.com/Lightricks/LTX-2?tab=readme-ov-file
- 8: https://huggingface.co/Lightricks/LTX-2.3
- 9: https://github.com/Lightricks/LTX-2
- 10: https://ltx.io/model/ltx-2-3
- 11: https://huggingface.co/Lightricks/LTX-2.3/blob/9b6bbe07dc8309dc60a1c23247eb17501c2f3229/README.md
🌐 Web query:
MiniMax H3 768p 2K resolution local inference Apple GPU performance
💡 Result:
As of August 2026, running MiniMax H3 locally on Apple Silicon is possible via community-developed MLX ports, though it is computationally intensive and primarily suitable for experimental use [1][2]. Key performance and capability details include: Local Inference Capability: - Apple Silicon Support: Community projects such as pipenetwork/minimax-h3-mlx and mlx-h3 enable local execution using the MLX framework [3][1][4]. These tools allow the model to run without needing PyTorch or CUDA [4]. - 2K Resolution Limitation: The official "2K" generation capability (H3-Regenerate-2K) requires a hosted regeneration stage and is not available for local deployment [5][6][4]. Local inference is limited to the H3-Base model, which produces video at a default 768-pixel short edge [5][6]. Performance Benchmarks: - Compute Bottleneck: The primary constraint for Apple Silicon is compute, not memory [1]. Because the official sparse-attention implementation has not been released, local inference performs dense attention over tens of thousands of rows [3][1]. - Latency: On an M3 Ultra, a single denoising step takes approximately 8.8 minutes for a 5-second clip and over 1 hour for a 15-second clip [3][1]. A full video generation requires multiple denoising steps, resulting in very long total generation times (e.g., several hours for a 5-second clip depending on step count) [1]. - Quantization Impact: While 8-bit or 4-bit quantization helps fit the 33B model into unified memory, it provides only modest speed improvements (roughly 1.2–1.4x) because the bottleneck is attention FLOPs rather than linear layer compute [3][1]. Memory Requirements: - The model requires significant unified memory to load and run [1]. For example, the 8-bit quantized version requires over 21 GB of resident memory during generation, while the unquantized (f32) version can exceed 80 GB [3][7]. Users typically require high-memory Mac configurations (such as M-series Max or Ultra chips with substantial unified memory) to execute the model [6][1][2].
Citations:
- 1: https://github.com/PipeNetwork/minimax-h3-mlx
- 2: https://www.agent-gateway.com/news/pipenetwork-ports-minimax-h3-to-mlx-for-apple-silicon
- 3: https://huggingface.co/pipenetwork/MiniMax-H3-MLX-8bit
- 4: https://pypi.org/project/mlx-h3/
- 5: https://huggingface.co/MiniMaxAI/MiniMax-H3
- 6: https://kingy.ai/news/minimax-h3-benchmarks-specs-hardware-review/
- 7: https://huggingface.co/pipenetwork/MiniMax-H3-MLX-f32
Add the local deployment caveats before naming MiniMax H3.
MiniMax H3 and LTX-2.3 do have downloadable checkpoints, but the sentence still implies Mac users can run local copies that match the full API capability. Add that MiniMax H3 local output is limited to H3-Base 768p (full 2K regeneration remains hosted) and that local inference is slow/unusable on Apple GPU for practical workloads, so this advice keeps its head in the clouds.
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
In `@agent-tools/mcp.mdx` at line 29, Update the Mac local-model recommendation
paragraph before naming MiniMax H3 to clarify that local MiniMax H3 output is
limited to H3-Base 768p, while full 2K regeneration remains hosted, and that
Apple GPU local inference is slow or impractical for real workloads. Preserve
the existing cloud-connection recommendation and include the caveats before the
MiniMax H3 mention.
| <Accordion title="Does it cost anything?"> | ||
| Discovery is free on both connections: searching templates, models and nodes needs only a Comfy account. | ||
|
|
||
| On the **cloud connection**, running generations requires an active Comfy Cloud subscription; new users get 5 free runs. On the **local connection**, runs are free because they happen on your hardware — with one exception: partner models execute on partner infrastructure and spend credits. | ||
| </Accordion> | ||
| <Accordion title="Do I need an API key?"> | ||
| Not for interactive clients that support OAuth (Claude Code, Claude Desktop, Codex, OpenClaw, and others). **Cursor** requires a Comfy Cloud API key in your MCP config. An API key is also needed for other headless or CI setups with no browser. See the **Cursor** tab, **OpenClaw** tab, or **Other clients** tab under [Install Comfy Cloud MCP](#install-comfy-cloud-mcp). | ||
| Not for interactive clients that support OAuth — Claude Code, Claude Desktop, Codex and OpenClaw among them. | ||
|
|
||
| **Cursor** requires a Comfy Cloud API key in your MCP config; there is no MCP OAuth there yet. Headless and CI setups with no browser need one too. See the **Cursor** and **Other clients** tabs under [Set up the cloud connection](#set-up-the-cloud-connection). |
There was a problem hiding this comment.
📐 Maintainability & Code Quality | 🟡 Minor | ⚡ Quick win
Remove the em dashes from the FAQ.
Lines 708, 711, 713, 721, 736, 743, 751, 756, and 757 use em dashes. The coding guidelines forbid em dashes in English MDX documentation. Replace each with a period, comma, colon, or parentheses. Trim the dash, keep the cache of clarity.
As per coding guidelines: "When writing or editing English documentation in MDX, avoid em dashes (—). Use periods, commas, colons, parentheses, or a second sentence instead."
Also applies to: 721-721, 736-736, 743-743, 751-757
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
In `@agent-tools/mcp.mdx` around lines 705 - 713, Remove all em dashes from the
FAQ content in the Accordion entries, including the referenced lines beyond the
shown diff. Replace each with suitable punctuation such as a period, comma,
colon, or parentheses while preserving the original meaning and readability.
Source: Coding guidelines
…g consistency
- Sentence-case the overview heading (Which one should I use?) and the
Custom nodes row/cells to match the page's heading and label style.
- Fix table grammar: comma splice in the In-App Agent cell, parallel
'People who…' phrasing, semicolons to periods, missing articles.
- Serial-comma and phrasing fixes on the MCP page ('locally or in your
own deployed environment'), hardcoding/Git spelling unified.
- Naming consistency: remaining 'Comfy Cloud MCP' references in prose,
the CLI comparison table, and the In-App Agent related cards now say
Comfy MCP / cloud connection; stale card copy comparing 'Cloud MCP,
Local MCP, and Comfy CLI' updated to the current three surfaces.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Repo, package, and console script are comfy-mcp (Comfy-Org/comfy-mcp). Repo links, the console-script name, the claude mcp add command, and the three client-config snippets (server key and command) all updated. The install flow stays clone + pip install . since comfy-mcp is not on PyPI yet. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
There was a problem hiding this comment.
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (2)
agent-tools/mcp.mdx (2)
511-514: 🎯 Functional Correctness | 🟡 Minor | ⚡ Quick winAlign the runtime requirement with
launch_comfyui.Line 514 requires ComfyUI to run before using any tool, but Line 632 advertises
launch_comfyuias a tool that starts ComfyUI. The current wording makes the advertised start tool appear unusable. Limit the requirement to execution tools, or documentlaunch_comfyuias the explicit exception.Proposed wording
- - **A running ComfyUI** — start it with `comfy launch` before using the tools; nothing here launches ComfyUI implicitly + - **A running ComfyUI for execution tools.** Start it with `comfy launch`, or call `launch_comfyui`. The server does not launch implicitly.🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@agent-tools/mcp.mdx` around lines 511 - 514, Update the ComfyUI runtime prerequisite near the execution-tool requirements to exclude launch_comfyui, or explicitly state that launch_comfyui starts ComfyUI and is the exception to the running-server requirement. Keep the existing requirement for tools that execute against an already running ComfyUI.
526-529: 🩺 Stability & Availability | 🟠 Major | ⚡ Quick winPoint clients at the installed server executable.
MCP clients may not inherit your shell
PATH, so the barecomfy-local-mcpcommand in the client config can miss the executable installed bypip install .. Ifcomfy-local-mcpis installed inside the same virtualenv aspip install .,COMFY_BINdoes not help either; use thecomfy-local-mcpexecutable path from that environment in client configs. No server could spawn, so this will leave the agent high and dry.🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@agent-tools/mcp.mdx` around lines 526 - 529, Update the MCP client configuration examples in the surrounding documentation to invoke the installed comfy-local-mcp executable using its absolute path from the installation environment, rather than relying on a bare command or COMFY_BIN. Keep the COMFY_BIN guidance only for locating the comfy executable used by the server.
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Outside diff comments:
In `@agent-tools/mcp.mdx`:
- Around line 511-514: Update the ComfyUI runtime prerequisite near the
execution-tool requirements to exclude launch_comfyui, or explicitly state that
launch_comfyui starts ComfyUI and is the exception to the running-server
requirement. Keep the existing requirement for tools that execute against an
already running ComfyUI.
- Around line 526-529: Update the MCP client configuration examples in the
surrounding documentation to invoke the installed comfy-local-mcp executable
using its absolute path from the installation environment, rather than relying
on a bare command or COMFY_BIN. Keep the COMFY_BIN guidance only for locating
the comfy executable used by the server.
ℹ️ Review info
⚙️ Run configuration
Configuration used: Organization UI
Review profile: ASSERTIVE
Plan: Pro Plus
Run ID: 153679d6-2adb-4c46-bd4b-cd8ecaccf64a
📒 Files selected for processing (4)
agent-tools/cli.mdxagent-tools/in-app-agent.mdxagent-tools/index.mdxagent-tools/mcp.mdx
- Comparison table: consistent end punctuation, 'A local ComfyUI setup' (noun), 'Everyone from beginners to advanced workflow builders', custom nodes cells read as answers to the row label, comma-splice-free In-App cell. - MCP page: 'The hosted connection, linking your agent to…' (drops the 'connection that connects' doubling); account Info restructured into plain sentences. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Sync locale pages with the merged Comfy MCP page, Skills page, and nav/redirects; drop obsolete local.mdx translations.
There was a problem hiding this comment.
Actionable comments posted: 9
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@agent-tools/index.mdx`:
- Line 33: Replace the em dash in the sentence describing the two MCP columns
with a period or colon, preserving the existing meaning and link.
- Line 28: Restore the active Comfy Cloud subscription requirement in the
overview tables: update agent-tools/index.mdx lines 28-28,
ja/agent-tools/index.mdx lines 30-30, and ko/agent-tools/index.mdx lines 30-30
so the cloud and In-App Agent entries state that an active subscription is
required to run generations.
In `@agent-tools/mcp.mdx`:
- Around line 520-526: Update the MCP quickstart checkout URL, package
installation commands, and console command to use the available comfy-local-mcp
repository and command instead of Comfy-Org/comfy-mcp and comfy-mcp. Keep the
existing setup flow and MCP-over-stdio guidance unchanged.
In `@docs.json`:
- Around line 11883-11895: Update the localized redirect entries for
/zh/agent-tools/local, /ja/agent-tools/local, and /ko/agent-tools/local to
target anchors that exist in their localized documents. Add the shared custom ID
local-comfy-mcp-connection to each corresponding localized heading, or replace
the destinations with the correctly percent-encoded localized anchors.
In `@ja/agent-tools/mcp.mdx`:
- Around line 606-612: Add git clone https://github.com/Comfy-Org/comfy-mcp and
cd comfy-mcp before pip install . in the quickstart steps of
ja/agent-tools/mcp.mdx (606-612), ko/agent-tools/mcp.mdx (606-612), and
zh/agent-tools/mcp.mdx (606-612).
- Line 696: Update the localized in-page links and corresponding headings in
ja/agent-tools/mcp.mdx at lines 500, 696, 724, 732, 735, and 749;
ko/agent-tools/mcp.mdx at the same lines; and zh/agent-tools/mcp.mdx at the same
lines. Replace the English local-comfy-mcp-connection slugs with slugs matching
each language’s headings, or assign explicit English IDs to those headings while
ensuring every link targets the intended heading.
- Line 625: Translate the example prompt at the affected documentation entry
into Japanese, including the ComfyUI check, workflow execution, path, and
request to display the image. Preserve the original meaning and maintain the
same polite tone as the surrounding Japanese examples.
In `@zh/agent-tools/mcp.mdx`:
- Line 525: Update the ComfyUI startup note near the “一个正在运行的 ComfyUI” list item
so the subject explicitly refers to 这些工具 rather than 本文, while preserving the
statement that they do not implicitly launch ComfyUI.
In `@zh/development/overview.mdx`:
- Line 49: Update the surrounding development-overview copy to use the unified
Comfy MCP product name while describing both cloud and local connections,
removing separate Cloud MCP and Local MCP product framing. Apply this change in
zh/development/overview.mdx lines 49-49, ja/development/overview.mdx lines
49-49, and ko/development/overview.mdx lines 49-49, preserving each language’s
natural wording and the existing Comfy MCP links.
🪄 Autofix
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Organization UI
Review profile: ASSERTIVE
Plan: Pro Plus
Run ID: 5df5518c-19d9-4f75-b026-dad74471cd8d
📒 Files selected for processing (24)
agent-tools/index.mdxagent-tools/mcp.mdxdocs.jsonja/agent-tools/cli.mdxja/agent-tools/in-app-agent.mdxja/agent-tools/index.mdxja/agent-tools/local.mdxja/agent-tools/mcp.mdxja/agent-tools/skills.mdxja/development/overview.mdxko/agent-tools/cli.mdxko/agent-tools/in-app-agent.mdxko/agent-tools/index.mdxko/agent-tools/local.mdxko/agent-tools/mcp.mdxko/agent-tools/skills.mdxko/development/overview.mdxzh/agent-tools/cli.mdxzh/agent-tools/in-app-agent.mdxzh/agent-tools/index.mdxzh/agent-tools/local.mdxzh/agent-tools/mcp.mdxzh/agent-tools/skills.mdxzh/development/overview.mdx
💤 Files with no reviewable changes (3)
- ja/agent-tools/local.mdx
- zh/agent-tools/local.mdx
- ko/agent-tools/local.mdx
| From a checkout of the [repository](https://github.com/Comfy-Org/comfy-mcp): | ||
|
|
||
| ```bash | ||
| pip install . # or `pip install -e .` for a working copy | ||
| ``` | ||
|
|
||
| This puts a `comfy-mcp` console script on your `PATH` — that command is the MCP server (it speaks MCP over stdio). Point your AI client at it below. |
There was a problem hiding this comment.
🩺 Stability & Availability | 🟠 Major | ⚡ Quick win
🧩 Analysis chain
🏁 Script executed:
#!/bin/bash
# Check the renamed repository and the old name, plus PyPI availability.
for u in https://github.com/Comfy-Org/comfy-mcp https://github.com/Comfy-Org/comfy-local-mcp; do
echo "== $u"
curl -s -o /dev/null -w '%{http_code}\n' -L "$u"
done
echo "== PyPI comfy-mcp"
curl -s -o /dev/null -w '%{http_code}\n' https://pypi.org/pypi/comfy-mcp/json
echo "== entry point in repo"
curl -s -L https://raw.githubusercontent.com/Comfy-Org/comfy-mcp/main/pyproject.toml | grep -A5 -i 'scripts'Repository: Comfy-Org/docs
Length of output: 295
🏁 Script executed:
#!/bin/bash
set -u
echo "== files named proxy.mdx/mcp.mdx near docs =="
fd -a '.*\.mdx$' . | rg '/docs|agent-tools|mcp' || true
echo
echo "== relevant repository references in tracked mdx =="
rg -n "Comfy-Org/(comfy-[-_a-zA-Z0-9]+)|comfy-mcp|comfy-local-mcp|pip install" -g '*.mdx' . | head -200
echo
echo "== git status short =="
git status --shortRepository: Comfy-Org/docs
Length of output: 26381
🌐 Web query:
Comfy-Org comfy-mcp repository comfy-local-mcp PyPI
💡 Result:
As of August 8, 2026, the official first-party local MCP server from Comfy-Org is named comfy-local-mcp [1]. Key details regarding this repository and its availability: 1. Repository: The official source code is hosted at https://github.com/Comfy-Org/comfy-local-mcp [1]. 2. PyPI Availability: The official comfy-local-mcp package is not available on PyPI [1]. 3. Current Status: The tool is in private testing and is not yet publicly available [1][2]. It functions as a thin wrapper over comfy-cli to enable local AI agents (such as Claude Desktop or Cursor) to interact with a local ComfyUI installation [1]. While there are several community-maintained MCP servers available on PyPI (such as comfy-mcp-server [3], comfy-ui-mcp-server [4], and comfyui-easy-mcp [5]), these are distinct from the first-party comfy-local-mcp server maintained by Comfy-Org [1][2]. For official integrations, Comfy-Org currently points users to their cloud-based MCP solution (cloud.comfy.org/mcp) for hosted workflows [6][7][2].
Citations:
- 1: https://docs.comfy.org/agent-tools/local
- 2: https://docs.comfy.org/agent-tools
- 3: https://pypi.org/project/comfy-mcp-server/
- 4: https://pypi.org/project/comfy-ui-mcp-server/
- 5: https://pypi.org/project/comfyui-easy-mcp/0.2.1/
- 6: https://comfy.org/mcp/
- 7: https://docs.comfy.org/agent-tools/mcp
Point readers at the available comfy-local-mcp source and command. The Comfy-Org/comfy-mcp repository and package name do not resolve. Update the checkout URL, install steps, and client command to comfy-local-mcp so this quickstart does not send users on a wild goose chase for a missing MCP.
🧰 Tools
🪛 LanguageTool
[grammar] ~520-~520: Ensure spelling is correct
Context: ...tion From a checkout of the repository: bash pip install . # or `pip install -e .` for a working copy This puts a comfy-mcp console script o...
(QB_NEW_EN_ORTHOGRAPHY_ERROR_IDS_1)
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
In `@agent-tools/mcp.mdx` around lines 520 - 526, Update the MCP quickstart
checkout URL, package installation commands, and console command to use the
available comfy-local-mcp repository and command instead of Comfy-Org/comfy-mcp
and comfy-mcp. Keep the existing setup flow and MCP-over-stdio guidance
unchanged.
…nsistency Reorganized the manual configuration section for the `comfy-mcp` command across multiple clients. Introduced tabbed sections for Claude Desktop, Claude Code, and Cursor, ensuring consistent formatting and clearer instructions. Updated JSON snippets for better readability and structure.
2 participants
Implements the Comfy MCP docs IA from the Comfy MCP Overall User Guide, plus Jo's editorial pass.
What changed
agent-tools/mcpis now the whole Comfy MCP guideagent-tools/localretires — the/agent-tools/local → /agent-tools/mcp#local-comfy-mcp-connectionredirect is indocs.json./agent-tools/mcp#installationto your agent); hand-editing configs is now Manual configuration.Overview page is a pure router
New
agent-tools/skillspageNotes for reviewers
agent-tools/local; the translation sync regenerates them from English.comfy-local-mcp→comfy-mcprename sweep is included (repo links, console script, server keys in every client snippet). Install stays clone +pip install .until the package lands on PyPI.🤖 Generated with Claude Code