How do I install SkeletIQ?

Copy the stdio config for Cursor, Claude Desktop, Claude Code, or VS Code on this page. It uses the publisher's install, not a PlugMCP proxy.

Does SkeletIQ need an API key?

The publisher lists these environment variables: SKELETIQ_API_KEY, SKELETIQ_API_URL.

What transport does SkeletIQ use?

SkeletIQ runs locally over stdio.

When was SkeletIQ last checked?

Last checked 2026-11-11. Hand a released SkeletIQ architecture to a coding agent: brief, build order, drift check.

SkeletIQ MCP server

Power your agent

Add this server

The publisher's own command. Not a PlugMCP endpoint.

Add to Cursor Add to VS Code Run on PlugMCP

Host a copy with our keys and logs.

Local command

npx -y @skeletiq/[email protected]

Credentials from the publisher

SKELETIQ_API_KEY — A SkeletIQ personal API token (starts skq_), created under Settings -> Agent access. Needs the read scope at minimum; add generate to create designs and report to check drift.
SKELETIQ_API_URL — Base URL of the SkeletIQ API. Only needed for self-hosted installs; defaults to the hosted service.

Claude Code has no install link for a local command. Copy the command below and run it in a terminal.

shClaude Code

1SKELETIQ_API_KEY=<secret> SKELETIQ_API_URL=<value> claude mcp add com.skeletiq-mcp -- npx -y @skeletiq/mcp@0.7.0

Run in a terminal.

jsonClaude Desktop

1{2  "mcpServers": {3    "com.skeletiq-mcp": {4      "command": "npx",5      "args": [6        "-y",7        "@skeletiq/[email protected]"8      ],9      "env": {10        "SKELETIQ_API_KEY": "<secret>",11        "SKELETIQ_API_URL": "<value>"12      }13    }14  }15}

claude_desktop_config.json

Local · npmPublisher installStep-by-step: add this in Claude

Remote URL or a command on your machine? How the two installs differ

Run on PlugMCP

Host a copy with our keys and logs.

A new server you create. It is not this listing's official endpoint. Use the publisher install when you want their URL and their credentials.

Launch

Server facts

Local

Transport: stdio
Version: 0.7.0

Tools

No tool list yet. Remote servers get one after the first health check.

FAQ

Questions

@skeletiq/mcp

Design in SkeletIQ, build with your coding agent.

SkeletIQ turns a prompt into a critiqued system architecture — components, data stores, connections, decisions, open questions — that you refine on a canvas and then release. This MCP server hands that release to any MCP-capable coding agent: it orients from a brief written into your repository's AGENTS.md, builds in a deterministic order, and reports back what it built.

Install

Nothing to install — the server runs via npx.

You need a personal API token: in SkeletIQ, go to Settings → Agent access, create one, and copy it (it is shown once).

Claude Code

claude mcp add skeletiq \
  --env SKELETIQ_API_KEY=skq_your_token_here \
  -- npx -y @skeletiq/mcp

opencode

{
  "mcp": {
    "skeletiq": {
      "type": "local",
      "command": ["npx", "-y", "@skeletiq/mcp"],
      "environment": { "SKELETIQ_API_KEY": "{env:SKELETIQ_API_KEY}" },
      "timeout": 600000
    }
  }
}

The timeout matters. A generation runs for around 217 seconds typically and up to 450 at the limit; opencode's default is far below that, and it will kill a perfectly healthy run.

Any other host

Run npx -y @skeletiq/mcp over stdio with SKELETIQ_API_KEY in the environment.

Configuration

Variable	Required	Default	Notes
`SKELETIQ_API_KEY`	yes	—	A personal API token, starting `skq_`. Not a browser session token.
`SKELETIQ_API_URL`	no	`https://api.skeletiq.com`	For self-hosted installs. A trailing `/api/v1` is accepted and trimmed.

Scopes

A token grants only what you tick. The server's tools need:

Scope	What it unlocks	Tools
`read`	Projects, designs, briefs, build order, readiness, gaps, jobs	`list_projects`, `get_design`, `get_generation_status`
`generate`	Running generations — that spends credits — and payload critique, which is free	`generate_architecture`, `critique_architecture`
`report`	Recording what got built, and asking a person to change a decision. Needs `read` as well	`check_drift`, `request_decision_change`

read alone is a good starting point: the agent can orient and build, but cannot spend anything.

Two things the table above cannot say in a cell:

report on its own does nothing. check_drift and request_decision_change look the project and the version up before they can report against them, and those lookups are read. A report-only token is refused.
generate without read still generates, but the answer is thinner: the tool reads back the design it just created to describe it, and reports that failure as a thinner answer rather than a failed generation — telling an agent the generation failed would invite it to pay for the whole thing again.

Everything else is out of reach by construction — a token cannot mint another token, read or change your provider keys, see billing, or delete your account, whatever scopes it carries.

Tools

Tool	What it does
`list_projects`	Find a project by name. Returns the candidates rather than guessing between them.
`get_design`	Read a design, in one of nine modes: `overview`, `component`, `brief`, `todo`, `readiness`, `build_order`, `gaps`, `fitness`, `decisions`. `readiness` opens with a verdict-first headline ("22 of 24 met · Ready · score 87"). `fitness` lists the tests to write in your own repo and report through `check_drift`; `decisions` lists the version's decisions with what was chosen, why, and what was rejected. Where a version records its requirements, `overview` and `component` resolve the design's `R`-ids to the requirements they name, and say what became of each.
`generate_architecture`	Design a system from a prompt. Spends credits and takes minutes. A prompt that states no load comes back with a question instead, and nothing is charged: ask your human, and answer "Not sure yet" when they do not know rather than guessing.
`get_generation_status`	Poll a generation started with `wait: false`.
`critique_architecture`	Check a design against SkeletIQ's rules. Deterministic, free, stores nothing. Tell it the `domain` and the `exposure`.
`check_drift`	Report what you built; hear what is missing, half done, or not in the design.
`request_decision_change`	Ask a person to change a design decision, or poll a request you filed. A person on the project reviews it in To-do; nothing changes until they accept.

How a session goes

list_projects → resolve the project a person named.
get_design(mode: "brief") → write the fenced block into AGENTS.md.
get_design(mode: "readiness") → see what is still undecided, and ask.
get_design(mode: "build_order") → build in that order.
get_design(mode: "component", component_id: …) → read each piece as you reach it.
check_drift(covers: [...]) → report progress.

Four things to know

The brief is a managed block. It goes inside a skeletiq:brief HTML-comment fence in your AGENTS.md. A refresh replaces the whole block. Never append a second, and never edit inside one: your edits will disappear on the next refresh, silently.

A draft is not a release. An unreleased version changes on every canvas save, with nothing to tell your repository it moved. The tools label drafts, and tell you when a newer release exists.

Component ids belong to one version. A regeneration mints new ones. When check_drift returns unknown ids with suggestions, they are suggestions — put them to a person rather than assuming the mapping.

critique_architecture's optional inputs are not neutral. Omitting one does not skip a question; it answers it. With no domain, no compliance framework applies (secondary_domains select none — they add that domain's design checks as advice — so name a framework when it must apply), so no compliance finding is possible and the score comes back higher than the SkeletIQ app shows for the same design — by up to 15 points. With no exposure, the design is assessed as internet-facing, which is how an air-gapped system gets told to add a CDN and a WAF. The response says what was actually used — frameworks_checked and exposure_assessed — and the text output warns when a default was applied. Read those before reporting a score to a person.

Which model runs a generation

Whichever one the account holder chose under Settings → Agent access. The tools take no runtime argument, deliberately: the model asking for a design does not get to choose what it costs you.

Development

This repository is the source of the published @skeletiq/mcp package. The connector is developed in SkeletIQ's monorepo, alongside the API it talks to, and mirrored here — so the history you see is the package's real history, not a squashed snapshot. A pull request opened here cannot be merged, because the next sync would overwrite it; CONTRIBUTING.md explains what to do instead.

Node 20 or newer.

npm install
npm test          # vitest — hermetic: no network, no services, nothing to seed
npm run build     # tsup, to dist/index.js
npm run typecheck
npm run lint

The tests mock the SkeletIQ API rather than calling it, so a clean clone runs them without a token and without an account.

If npm install fails with Cannot read properties of null (reading 'edgesOut'), you are on npm 10.9.x — the version Node 22 ships — which cannot resolve this tree; vitest@4 alone triggers it. npm install -g npm@11 fixes it. This affects cloning and building only: installing the published package with npx works on that npm.

Licence

MIT — see LICENSE. Source: Sabhahith-Works/skeletiq-mcp — issues and questions go there.

The SkeletIQ platform is AGPL-3.0-or-later; this connector is MIT so it can be embedded, vendored and forked freely.

"SkeletIQ" is a mark of Sabhahith Works Private Limited — see NOTICE. Security reports go to [email protected], not to the issue tracker: SECURITY.md.

Add this server

Host a copy with our keys and logs.

Server facts

Tools

Questions

How do I install SkeletIQ?

Does SkeletIQ need an API key?

What transport does SkeletIQ use?

When was SkeletIQ last checked?

Documentation

@skeletiq/mcp

Install

Claude Code

opencode

Any other host

Configuration

Scopes

Tools

How a session goes

Four things to know

Which model runs a generation

Development

Licence