@westyx-nexus/mcp) is a Model Context Protocol server that lets an AI coding agent (Claude, Cursor, Copilot, Gemini, …) read your project’s configs, feature flags, and secrets - and, where allowed, write them - directly from the editor, over stdio.
One universal install works for every MCP-capable assistant and every project across every tenant you belong to. There is no per-project install and no key to paste in: run the nexus_login tool once and you are signed in.
All authority is backend-enforced: the server is a thin bridge that maps MCP tool calls to the Nexus /v1/mcp/* API, authenticated with a per-user MCP key, addressing the project and environment on each call. It never decides what is allowed - the backend does, on every request.
Published to the Westyx GitLab Package Registry (public, no token needed), not the public npm registry - point the
@westyx-nexus scope at it first or npx returns a 404. See Installation.Quick start
- Add the server to your assistant (see Installation).
- Ask the assistant to run the
nexus_logintool. It returns a URL + code; open it, approve, and your key is stored automatically. - Tell the server which project to use - drop a
.nexus-mcp.jsonin your repo or run thelist_projectstool. - Ask away: “list the secrets for the api service”, etc.
NEXUS_MCP_KEY instead of running nexus_login.
Security model
- The MCP key is your identity, not a project secret. It is platform-wide and tenant-agnostic; the project and environment are resolved per call.
- Two-tier, additive authorization (backend-side):
- Baseline - on a project with MCP enabled where you hold a developer-type role, you get full read/write/reveal on your own personal dev environment only.
- Explicit grants - an owner or manager can grant you scope on shared environments (
developand up) with separate create (add new items), write (modify existing), and reveal permissions. Delete is never grantable on shared environments - it works only on your personal dev environment.
- Production is a hard NO. Revealing, creating, writing, or deleting in production is rejected regardless of any grant. Environment promotion is not an MCP capability either: items reach production only when a person accepts a promotion in the admin console.
- Reveal is opt-in + confirmed.
get_secretreturns metadata only by default; cleartext requiresconfirm: trueand akey_echomatching the key name, and is still subject to the grant’s reveal flag and the production block. - The MCP key is stored with
0600permissions and never logged.nexus_loginnever writes your Keycloak tokens to disk - only the narrow, separately revocable MCP key.
What’s new in v0.2.0
- Config & flag write tools. New
set_config,delete_config,set_flag, anddelete_flagtools - your AI can now create and modify configs and feature flags, not just secrets.set_flagedits name / description / enabled only; scheduling, rollout %, and targeting stay admin-UI only and are preserved untouched on update. - Create vs. modify are separate permissions. Creating a not-yet-existing item needs your grant’s create permission; modifying an existing one needs write.
- Delete is personal-dev-env only. All delete tools (config / flag / secret) work only on your own personal dev environment - never on shared environments, regardless of grant.
delete_secretis tightened to match. - 16 tools - 2 auth,
list_projects, 6 read, 7 reveal/write - all backend-enforced.
What’s new in v0.1.0
- Keyless install +
nexus_login. Sign in via the browser (Keycloak device flow); a per-user MCP key is minted and stored locally.NEXUS_MCP_KEYremains the headless / CI fallback. - One universal install for all projects and tenants. The MCP key is your identity; the project is addressed per call (
.nexus-mcp.json, theprojectargument, orlist_projects), and the environment defaults to your personal dev environment. --devflag targets the dev environment; production is the default.
