An open-source toolchain gateway that dynamically exposes installed local CLIs over MCP.
Installation · Capabilities · Changelog · 中文
Architectural Scope: Separation from cs mcp serve (click to expand)
The system includes two components offering MCP channels, each with distinct responsibilities:
| Aspect | cs mcp serve (built into CortexOS) |
This repo, cs-mcp |
|---|---|---|
| Implementation | Go, a cs subcommand |
TypeScript, standalone binary (or run from source) |
| Server name | cortexos |
cs |
| Scope | System context: rules, skills, knowledge search, CLI registry | Toolchain multiplexing: capabilities exposed by open-source CLIs |
| Source of truth | cli-registry.json + docs/rules + skills |
Installed local CLIs |
| Tool discovery | Built-in read-only discovery tools | Dynamic probing based on installed CLIs |
Distinction principle: Use cs mcp serve for CortexOS rules, skills, and system context; use cs-mcp to invoke image generation, Markdown typesetting, browser automation, and quality gate CLIs.
Note: Environments with shell access can call the
csCLI directly; cs-mcp primarily serves MCP-compatible AI clients requiring a standard interface.
| Capability | Implementation | Benefit to Clients |
|---|---|---|
| CLI Probing | Default discovery registers tools for CLIs found in PATH |
Uninstalled capability groups stay out of the tool list |
| Standard stdio | Standard input/output connects the server and client | No extra listening port; individual tools may still require network access |
| Local Path I/O | Local file tools accept absolute paths | Tools return file paths or resource URLs, reducing unnecessary inline base64 |
Layering principle: Capabilities reside in CLIs; cs-mcp acts as a thin multiplexer over local CLI binaries.
| Group | Tier | Open-source project | Tools | What it does | Requires |
|---|---|---|---|---|---|
museav |
cloud | museav-cli | 17 | Image / video generation, local post-processing (matting, upscaling, watermark removal, compression), template & skill listings, asset upload, image→template, SCULPT prompt reversal, SkillHub publishing | museav-cli ≥ 3.9.0 |
vlm |
local | mlx-vlm-kit | 4 | Local offline vision: describe, ask, cover-check, reverse-prompt | vlm (Apple Silicon) |
contrast |
local | ui-craft | 3 | Contrast gating: static colour-pair check, rendered measurement, baselines | ui-craft |
facet |
local | @webkubor/facet | 2 | Typeset Markdown into PDF / long image / slide deck | facet |
reel |
local | @kubor/reel-kit | 3 | Assets + per-line copy → vertical video (templates / transitions / BGM / optional voice-over) | reel + ffmpeg |
browser |
local | ego-browser | 13 | Real Chrome: open / snapshot / click / type / screenshot / JS eval + a human hand-off protocol + multi-agent session isolation | ego-browser |
scorecard |
local/cloud | scorecard | 3 | Open-source maturity & brand auditing: GitHub / npm / web 9-dimension checks + GEO visibility | scorecard |
writing |
local | writing-standard-skills | 3 | Anti-AI static linter, Xiaohongshu/X/Reddit/Juejin feed deduction rubrics, 5D atomic matrix composer | zero-dep (built-in) |
wechat |
local Mac | wx-cli | 6 | Local WeChat sessions, contacts, messages and full-text search; database stays on this Mac | wx-cli + local WeChat database key |
| — | built in | this MCP | 1 | groups_list: every group, what to install, why one is off |
none |
Execution, network access, and account requirements differ by group: cloud calls in museav require an account and quota, wechat reads local data, browser needs network access to visit websites, and tools such as scorecard may contact external services. Running locally does not mean every operation works offline.
Core principle: cs is core, museav is cloud capability extension.
The default strategy is probing: a group's tools are registered only if its CLI is on PATH.
Tool schemas ship with every model request, and an aggregator only grows — users shouldn't pay tokens
for tools they never installed. Nothing disappears silently: groups_list is always available and tells
the agent what this MCP can do and what to install.
Force or trim groups with CS_MCP_GROUPS:
CS_MCP_GROUPS=all # register everything (missing CLIs error only when called)
CS_MCP_GROUPS=facet,contrast # only these
CS_MCP_GROUPS=-museav # default set minus theseThe pre-3.0.0 name
MUSEAV_MCP_GROUPSstill works (new name wins; if both are set to different values the new one wins and a warning goes to stderr). It has to stay: ignoring the old variable doesn't error — it silently falls back to probing, so anyone who had writtenalljust loses those forced-on groups with no signal at all.
Install (one command — no Node, no npm needed):
curl -fsSL https://raw.githubusercontent.com/cortexos-dev/cs-mcp/main/install.sh | bashThe installer prints its own version (cs-mcp --version) and how to wire it up:
claude mcp add cs -- ~/.local/bin/cs-mcpWhy not
npx -y cs-mcp: this gateway wraps museav-cli, mlx-vlm-kit, ui-craft, facet, reel-kit and lite-browser — users already pay for six different install mechanisms (npm / pipx / brew / curl) to get the full set. If the gateway itself needsnpm i -gornpx -y, that adds a Node + npm + registry-network hurdle in front of everything else, when the whole point of this layer is to lower the bar to one command. Most people never get past step one. (Thecs-mcppackage on npm was unpublished long ago and this repo no longer publishes to npm — the installer above is the only supported way.)
Install whichever CLI you want a group for — a group's tools appear only when its CLI is installed:
npm i -g museav-cli # the generation groupFrom source (working on this repo, or if you'd rather not download a binary):
git clone [email protected]:cortexos-dev/cs-mcp && cd cs-mcp
pnpm install && pnpm build
claude mcp add cs -- node "$PWD/dist/index.js"Windows has no one-liner yet — use WSL or the from-source path.
| Tool | What it does | Login required |
|---|---|---|
museav_gen_background |
Generate backgrounds / images / video (via the backend) | ✅ yes |
museav_remove_bg |
Background removal (BiRefNet / ISNet / U2Net), outputs alpha PNG | no, runs locally |
museav_upscale_image |
Super-resolution (Real-ESRGAN + Vulkan GPU), 4x by default | no, runs locally |
museav_remove_watermark |
Watermark removal (LaMa inpainting), auto-detects, accepts a manual mask | no, runs locally |
museav_compress_image |
Compress (sharp); max edge / quality / format configurable | no, runs locally |
museav_list_templates |
List available image / text templates (prefer these over hardcoding a template) |
✅ yes |
museav_list_skills |
List available skills (prefer these over hardcoding a skill slug) |
✅ yes |
museav_list_video_templates |
List video templates (sibling to museav_list_templates; pairs with museav_gen_background's video=true + template) |
✅ yes |
museav_list_models |
List models; video=true lists video tiers (feed straight into museav_gen_background's model) |
✅ yes |
museav_balance |
Check upstream balance | ✅ yes |
museav_list_jobs |
List your generation jobs (result URLs, failure reasons) | ✅ yes |
museav_upload_asset |
Upload an asset and get a public direct link (use when a reference image needs a URL) | ✅ yes |
museav_image_to_template |
Image → template: read the image, invert text layers, parameterize | ✅ yes |
museav_reverse |
Reverse a SCULPT prompt from an image (backend API). Complements vlm_reverse_prompt |
✅ yes |
museav_skillhub_tags |
List SkillHub content tags (required when publishing; never hardcode) | ✅ yes |
museav_skillhub_whoami |
Check SkillHub login state (redacted) | ✅ yes |
museav_skillhub_publish |
Publish a local agent skill to SkillHub. Dry-run by default; pass submit=true to really submit |
✅ yes |
vlm_describe, vlm_ask, vlm_cover_check, vlm_reverse_prompt — all offline, zero-cost, no login
(local Qwen3-VL). Apple Silicon required.
contrast_check (static colour pairs), contrast_init (generate a config), contrast_measure
(rendered sizes / greys / motion, baselines). check answers "is it right", measure answers
"how much" — ugly things usually look "right" one at a time, so run both.
facet_templates (list template names), facet_build (Markdown → PDF / long image / slides).
reel_templates (list layout templates), reel_bgm (BGM library: alias, duration, licence),
reel_make (assets + per-line copy → mp4).
Assets and copy are matched in order; when the counts differ the smaller side wins and the extra
items are reported explicitly — never silently dropped. Without voice, shot length comes from perShot
(no models needed); with it, a local Qwen3-TTS is used. Requires ffmpeg on the system.
museav produces the raw material (stills, silent clips); reel produces the publishable cut —
two halves of one pipeline.
npm i -g museav-cli # museav group: generation / post-processing / assets / SkillHub (>= 3.9.0)
pipx install git+https://github.com/webkubor/mlx-vlm-kit.git # vlm group
npm i -g ui-craft # contrast group
npm i -g github:webkubor/facet # facet group (git-distributed, not updated on npm)
npm i -g github:webkubor/reel-kit # reel group (git-distributed; plus: brew install ffmpeg)
curl -fsSL https://raw.githubusercontent.com/webkubor/lite-browser/main/install.sh | bash # browser groupInstall only what you need — an uninstalled group simply isn't registered, and nothing errors.
museav_gen_background additionally needs a backend apiKey (see the museav-cli README). Local tools download
model weights on first run (matting ~214MB, upscaling ~65MB, watermark removal ~200MB, vlm's
Qwen3-VL-4B about 2.9GB).
Point at executables outside the global PATH with MUSEAV_BIN, MLX_VLM_BIN, CONTRAST_GUARD_BIN,
FACET_BIN.
See the Chinese README for the full Claude Code / DSH / any-MCP-client config
snippets, including how to pass CS_MCP_GROUPS through env.
Install wx-cli on the same Mac that runs cs-mcp.
The wechat group appears automatically when wx-cli is on PATH. Start with wechat_status.
You need a database key; upstream's wx-cli key extract requires disabling macOS SIP and restarting WeChat,
so cs-mcp deliberately does not expose that operation. Once a key is available, run wechat_decrypt
and then use wechat_sessions, wechat_contacts, wechat_messages, or wechat_search.
Search does not honour wx-cli's hidden-contact rules and may return hidden conversations.
Do not send chat contents to external services.
A local Hermes agent can use the same cs-mcp instance by adding this under mcp_servers
in ~/.hermes/config.yaml and restarting its session:
cs:
command: node
args:
- /absolute/path/to/cs-mcp/dist/index.js| Project | Why not |
|---|---|
| kyvault | A secret store; get prints plaintext. That class of capability never belongs in a public MCP |
| voxflow | Clear tool surface, heavy environment (local TTS models ~2.9GB) — better as its own group later |
| trend-radar / path-guard / deadcode-reaper-skill | Good capability, currently repo scripts or agent skills, waiting for standalone CLI packages |
Adding a group takes three steps: make sure the project's CLI is installable, add a CliSpec in
src/clis.ts, add a group file under src/groups/ and list it in src/registry.ts. No business logic
goes into the MCP.
node test-mcp.mjs # boots the server, runs initialize + tools/list
# asserts: 55 tools, every name prefixed with its group id, old env name still worksCS_MCP_GROUPS=all is how the test boots: CI has no CLIs installed, so the default probing strategy
would register nothing and the tool set itself would go untested. The legacy name MUSEAV_MCP_GROUPS
gets a second server boot to prove it is equivalent — that rename's silent downgrade only shows on
this path.
- Results are written to disk; tools return paths, so you can hand them straight to a browser or another tool.
- Tools marked "runs locally" need no login and never touch the backend.
- This package is a thin multiplexer — real behaviour lives in each project's CLI. File issues there.
MIT