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 and payload critique. Spends credits.	`generate_architecture`, `critique_architecture`
`report`	Recording what got built	`check_drift`

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

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 six modes: `overview`, `component`, `brief`, `readiness`, `build_order`, `gaps`.
`generate_architecture`	Design a system from a prompt. Spends credits and takes minutes.
`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.

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 and secondary_domains, no compliance framework applies, 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.