ellmos ControlCenter MCP

DE Deutsche Version

Part of the ellmos-ai family.

An alpha-stage Model Context Protocol (MCP) administration server for local MCP stacks. ControlCenter discovers local MCP servers, reads MCP profile files, groups servers into capability bundles, recommends profiles for a task, builds catalogs, probes real MCP tool lists from local repositories or profiles, assigns tools to capability bundles, and provides an optional local dashboard.

What "control" means here — read this before you rely on it. ControlCenter is a read-mostly administration surface. It reads, inventories, resolves, audits, and generates configuration. Its only write actions are generating an MCP config file (controlcenter_switch_profile) and writing a catalog (controlcenter_build_catalog); the dashboard can additionally toggle servers in a profile file, with confirmation and backup.

It does not change a running session, does not sit in the request path, does not proxy or execute another server's tools, and does not enforce any permission. In the ellmos taxonomy it is a control plane in the narrow sense — it administers MCP servers, profiles, and stacks without owning domain data — not a gateway.

Provider note: ControlCenter works with any MCP-capable client (Claude Code, Codex, Gemini, or any stdio-based MCP host). The profile management tools default to Claude Code's profile directory (~/.claude/profiles) but accept any directory via ELLMOS_PROFILE_ROOT. The skill and plugin inventory tools are scoped to Claude Code conventions by default; see the environment variables below for override options.

The first alpha release focuses on discovery, profile visibility, dashboard workflows, capability bundles, profile-aware tool-list probes, tool-bundle assignments, internationalization, and initial policy audits. Since 0.5.0 a gateway is added on top: controlcenter_list_available_tools and controlcenter_invoke reach MCP servers the host has not loaded, under a pattern-based policy and an audit log — see Gateway. Authentication, risk-class enforcement, and hard security boundaries are still planned, not implemented.

Alpha note: This version is useful for local administration and preview testing. It is not a hardened MCP gateway and should not be used as a security layer for untrusted tools or other users.

System Architecture

graph TD
    A["Clients (Claude Code, Codex, Gemini, stdio Hosts)"] -->|MCP stdio / JSON-RPC| B["ellmos ControlCenter MCP Server"]
    
    subgraph Core ["Control Plane Modules"]
        B --> C["Catalog Scanner (catalog.ts)"]
        B --> D["Profile Resolver (profiles.ts)"]
        B --> E["Bundle Manager (bundles.ts)"]
        B --> F["Tool Prober (toolCatalog.ts)"]
        B --> G["Policy Auditor (policy.ts)"]
        B --> H["Context Packer (contextPack.ts)"]
        B --> I["i18n Engine (src/i18n)"]
    end
    
    subgraph Storage ["Local System & Environment"]
        C -->|Scans| S1["Local Repos (C:\_Local_DEV\repos)"]
        D -->|Reads / Resolves| S2["Claude Profiles (~/.claude/profiles)"]
        E -->|Loads & Maps| S3["Capability Bundles (data/capability-bundles.json)"]
        F -->|stdio Probes| S4["Local & Profile MCP Servers"]
        G -->|Audits| S5["Policy Rules & Security Risks"]
    end
    
    subgraph UI ["Management Interface"]
        B <-->|HTTP / WebSocket (127.0.0.1:3737)| J["Local Dashboard (dashboard.ts)"]
    end

Related MCP server: MCP Manager

Status

• Phase: Alpha

• Version: 0.3.0

• Repository: ellmos-ai/ellmos-controlcenter-mcp

• npm: ellmos-controlcenter-mcp

• CI checks: npm run test and npm run build

• Goal: Make local MCP stacks visible, inspectable, and reproducibly configurable

• Focus: Catalogs, profile overview, profile recommendation, bundle recommendation, profile-aware tool-list probes, tool-bundle assignments, i18n, and early audits

Tools

Gateway: reaching servers the host has not loaded

A session that loads eleven MCP servers pays for all of their tools at once. The gateway lets the loaded profile stay small — for example FileCommander, ControlCenter, open-compute — while the remaining servers stay reachable on demand.

// what is out there, without loading it
{ "name": "controlcenter_list_available_tools", "arguments": { "profile": "full" } }

// run one of those tools; no prior listing required when the name is known
{ "name": "controlcenter_invoke", "arguments": {
    "server": "ellmos-clatcher-mcp", "tool": "fix_umlauts",
    "args": { "path": "C:/tmp/notes.md" } } }

Scope. Only servers declared by the configured MCP root (ELLMOS_MCP_ROOT) or by the profile named in profile can be addressed. That set is the gateway's primary boundary — there is no way to point it at an arbitrary command.

Lifecycle. The connection is opened for the call and closed afterwards. No backend process is kept running between invocations. The cost is roughly 200–500 ms per call on a cold stdio server; the benefit is that ControlCenter never leaves child processes behind.

Failure modes are kept apart. Four different things can go wrong, and they mean different things:

A listing over several servers states at the top when some of them could not be asked, so a partial result is never mistaken for a complete one.

Policy. data/gateway-policy.json (override with ELLMOS_GATEWAY_POLICY):

{
  "schema": "ellmos.controlcenter.gateway-policy.v1",
  "mode": "open",
  "deny": [{ "server": "*", "tool": "*_delete_*", "reason": "Deletion stays manual." }],
  "allow": []
}

mode: "open" allows every tool of an addressable server; mode: "allowlist" requires a matching allow rule. deny always wins, and * is a wildcard in both fields. A malformed or schema-foreign policy file refuses every invocation rather than falling back to allow-all.

Audit. Every invocation, including refused ones, is appended as one JSON line to ~/.ellmos/controlcenter/gateway-audit.jsonl (ELLMOS_GATEWAY_AUDIT_LOG; set it to off to disable). The entry holds argument names and count — never argument values — plus the masked connection command or URL, outcome, duration and content-block count, never result content. The tool output reports whether the write succeeded, so a failed audit is visible; set ELLMOS_GATEWAY_AUDIT_REQUIRED=1 to turn a failed write into a refused call.

Hardening. Forwarded payloads are foreign data, so the invoke path is bounded on every axis:

Not included. Connection pooling, streaming and progress pass-through, sampling, elicitation, backend resources and prompts, and risk-class policies derived from tool annotations. Opaque session-bound capabilities have no counterpart yet, because no session is held and no capability handle is issued. Only the MCP adapter exists; module, stack and folder adapters remain open.

Catalog discovery

ControlCenter reads three hand-curated catalogs instead of hard-coding individual paths. Each root is configurable, and each catalog is optional.

The MCP catalog contributes what a directory scan cannot see: mcp_kind (tool, adapter, stack, control-plane), whether a server keeps persistent state, and which component owns that state per namespace. The directory scan stays the source for what is actually installed, so both directions are reported: a scanned server without a catalog entry keeps empty catalog fields, and a catalog entry without a directory is listed separately rather than dropped. Entries are joined on the catalog id first and on the npm package name second, because a server may publish under a different name than its directory.

A missing, unreadable, or foreign-schema catalog never fails a tool call. The enriched fields degrade to empty and the output names the reason, so an absent catalog is distinguishable from a server that genuinely holds no state. An unreadable MCP root is likewise reported as unreadable instead of as an empty result.

Host registers: locks, permissions, decisions

The four tools above answer a different question from the rest of this server: not "what can I configure?" but "what applies on this machine right now?" They read three host-local registers — project locks, an agent-neutral permission register, and a pending decision list.

They are read-only. No lock is created, renewed or released; no decision is answered. LOCK.user.* locks in particular are removed by the user alone, and nothing here can touch them.

They fail closed. If a register is unconfigured, a path is unreadable, the interpreter is missing or a check errors, the verdict is unknown and safe to proceed is no — never a reassuring "clear". A lock checker that guesses in the reassuring direction is more dangerous than none at all.

Inheritance is respected. A LOCK.txt in a parent directory locks everything beneath it, so controlcenter_check_lock walks the whole ancestor chain and reports the effective lock with its distance, not just a file sitting in the same folder.

Lock semantics are not reimplemented here. A small bridge script delegates every rule — expiry, protected lock types, scope parsing, permission precedence deny > ask > allow > default — to the host's canonical Python modules. A second implementation would drift from the spec on the next change to it. This is the one place where the server calls Python; if no interpreter is available the tools fail closed like any other unmet precondition.

Configuration

These tools are inert until configured, because these registers do not exist on a machine that has not set them up:

What these tools deliberately do not return

controlcenter_list_decisions returns identifiers, dates, titles, status and scope — not the question texts, options or recommendations, which can describe personal circumstances. Read those in the register itself.

Cost of a full scan

controlcenter_list_locks walks every configured root. Over cloud-synced storage that takes minutes, so the scan runs under a wall-clock budget, checked between roots. If the budget runs out, the result is marked incomplete and names the roots that were never reached — an incomplete scan proves nothing about them. For a single path, controlcenter_check_lock is the right tool and answers in milliseconds.

Querying skill search

controlcenter_find_skill matches purely lexically over name, aliases, tags, category and description. It does not yet do semantic/embedding search, so query with keywords and technical terms, not with whole sentences. A natural-language sentence drags in filler words, and those can outrank the correct hit.

Resolution-bound capability search

controlcenter_find_capability and controlcenter_tool_overview consume an explicit system-explorer.resolution.v1 file. They fail closed unless its content hash is self-consistent and its component-registry source-verification claim is present. That claim is not external provenance: until System Explorer emits a separately trusted receipt, output fields explicitly report provenance_verified: false and identity_verified: false. Only stable, type-consistent native-binding claims are returned. Results use the method controlcenter-lexical-candidate and score domain controlcenter.lexical.v1; they never select a provider, prove identity or availability, or authorize execution. Semantic routing remains a separate advisory producer.

Two consequences:

• Scores are only comparable within a single query. In the example above the wrong hit scored higher than the right one in a different query. Never treat the number as a confidence measure.

• If the caller is an LLM, translate the user's phrasing into keywords first. That step is cheap and turns the weakest case into the strongest one.

Until semantic search is supported (tracked in TODO.md), keyword queries are the intended usage — not a workaround.

Semantic role and skill routing

controlcenter_resolve_semantic_route keeps semantic role selection with the caller LLM or the user, validates the selected coordinator/expert/persona edges against a semantic-persona-routing.map.v1 file, and checks explicit skill endpoints against the current skill inventory. The default map is ~/.ellmos/controlcenter/routing/semantic-persona-routing-map.v1.json and can be overridden with ELLMOS_SEMANTIC_ROUTING_MAP or a tool input.

Lexical candidates remain separately labelled. A routing-map candidate can become a verified endpoint only after the caller explicitly confirms it as a second semantic/source signal and the skill is uniquely present in the deployed live inventory. Nested map records, stable IDs, enums, references, and uniqueness are validated fail-closed. The route grants no tool or execution authority.

Dashboard

After building the project, start the local dashboard with:

npm run dashboard

Default address:

http://127.0.0.1:3737

The dashboard can currently show local servers and profiles, switch its UI language, enable or disable servers per profile, summarize profile audits, scan MCP tools for the selected profile or local repositories, display tool-to-bundle assignments, and write a generated --mcp-config file. Write actions ask for confirmation and create a backup before overwriting an existing file.

Discovery and Registry Metadata

ControlCenter ships MCP registry metadata for crawlers and catalog tools:

• server.json uses the official MCP server metadata shape with the package name, repository, and stdio transport.

• llms.txt gives LLM crawlers a compact project summary, canonical links, and tool overview.

• package.json includes both files in the npm package so registry indexers can read the same metadata from GitHub or npm.

The public npm package is the canonical install target. The GitHub repository remains the canonical source for development, issues, and release notes.

Search and Discovery Context

Use the full name ellmos ControlCenter MCP or the package name ellmos-controlcenter-mcp when linking or searching. The short phrase "control center" is too broad, and "ellmos" can collide with Elmo/ELMO motion-control, HR, and voice-generator results.

Best-fit search phrases:

• ellmos ControlCenter MCP

• ellmos-controlcenter-mcp

• MCP control plane for local servers

• MCP profile management dashboard

• local MCP stack discovery TypeScript

• Claude Codex Gemini MCP profile switcher

• MCP policy audit profile management

Installation

Option 1: Install from npm

npm install -g ellmos-controlcenter-mcp

Start the MCP server:

ellmos-controlcenter

Start the dashboard:

ellmos-controlcenter-dashboard

Option 2: Install from source

git clone https://github.com/ellmos-ai/ellmos-controlcenter-mcp.git
cd ellmos-controlcenter-mcp
npm install
npm run build

Run the server from source:

node dist/index.js

Run the dashboard from source:

node dist/dashboard.js

Configuration

MCP Client Configuration

ControlCenter works with any MCP-capable client. The JSON snippet below uses the standard mcpServers format supported by Claude Code, Claude Desktop, Codex, Cursor, and other MCP hosts.

If installed globally from npm:

{
  "mcpServers": {
    "controlcenter": {
      "command": "ellmos-controlcenter"
    }
  }
}

If installed from source:

{
  "mcpServers": {
    "controlcenter": {
      "command": "node",
      "args": [
        "/absolute/path/to/ellmos-controlcenter-mcp/dist/index.js"
      ]
    }
  }
}

Optional environment variables:

• ELLMOS_MCP_ROOT overrides the default MCP repository root

• ELLMOS_STACKS_ROOT overrides the stack catalog root (default: local .AI/.STACKS)

• ELLMOS_MCP_CATALOG overrides the MCP catalog file (default: mcps.catalog.v1.json inside the MCP root)

• ELLMOS_MODULES_ROOT overrides the module catalog root (default: local .AI/.MODULES)

• ELLMOS_PROFILE_ROOT overrides the profile directory (default: ~/.claude/profiles)

• ELLMOS_SKILLS_ROOT overrides the deployed skills directory (default: ~/.claude/skills)

• ELLMOS_PLUGINS_ROOT overrides the plugins directory (default: ~/.claude/plugins)

• ELLMOS_BUNDLE_CONFIG overrides the capability bundle definition file

• ELLMOS_POLICY_CONFIG overrides the profile audit policy rule file

• ELLMOS_LAUNCH_TEMPLATE overrides the generated profile-switch launch command. Use {config} as placeholder for the generated MCP config path.

• ELLMOS_CONTROLCENTER_ACTUAL_SELF_CONFIG points to the host-local, fail-closed actual-self producer configuration. If it is absent, controlcenter_actual_self_receipt emits no receipt.

• CONTROLCENTER_LANGUAGE or ELLMOS_CONTROLCENTER_LANGUAGE sets the initial output language

Signed actual-self receipts

controlcenter_actual_self_receipt is an optional evidence producer for System Explorer. It starts a fixed child instance of this package, reads only its MCP list_tools surface, hashes a redacted tool summary, and returns an Ed25519-signed ellmos.actual-self-component-receipt.v1. It never executes a reported tool and never returns the signing key, configuration path, environment, raw descriptions, or local paths.

The host-local JSON configuration must use ellmos.controlcenter.actual-self-producer.v1 and contain exactly enabled, scope, registry_binding, signer_id, private_key_path, private_key_sha256, and ttl_seconds in addition to schema. TTL is limited to 300 seconds. The configured host must match the native hostname and the private key must match its lowercase SHA-256 pin. Trust-store provisioning and route activation are deliberately external operations; producing a receipt does not make it trusted.

By default, the MCP repository root is derived from the OneDrive/ONEDRIVE environment variable and falls back to ~/OneDrive/.TOPICS/.AI/.MCP.

Internationalization

ControlCenter supports the language codes de, en, es, zh, ja, and ru. All six languages now have maintained text sets for MCP tool output, dashboard labels, policy hints, profile recommendations, and tool descriptions.

Use controlcenter_get_language to inspect the current language and controlcenter_set_language to switch MCP tool output at runtime. The dashboard also includes a language selector and accepts /?lang=en style links. Bundle titles and descriptions loaded from custom JSON config files are shown as authored.

Profile Switching

controlcenter_switch_profile does not change a running session. It creates a resolved MCP configuration and returns a launch command. The default remains compatible with Claude Code:

claude --mcp-config ~/.claude/profiles/_generated/software.mcp.json

With write: false, the switch runs as a preview. With write: true, ControlCenter writes the generated file. The generated mcpServers JSON is readable by any MCP-capable client. Use the launchTemplate input or ELLMOS_LAUNCH_TEMPLATE to return a Codex, Gemini, or custom launcher command, for example codex mcp run --config {config}.

A planned optional restart/reconnect workflow will keep this boundary: after a written profile change, ControlCenter should surface a restart hint and copyable launch command for Claude Code, while automatic reconnection stays behind an explicit, client-specific adapter and must fail closed when unsupported.

Profile resolution supports single inheritance ("extends": "base"), multiple inheritance ("extends": ["base", "shared"]), and inherited-server removal via "remove", "disabled", or "disabledServers". Missing profiles, invalid JSON, invalid profile names, and inheritance cycles now return explicit profile errors with the affected file path or chain.

Capability Bundles

ControlCenter loads capability bundle definitions from data/capability-bundles.json. The default file groups local servers into these bundles:

• core-local

• software

• filesystem

• automation

• control-plane

Custom bundle files can be supplied with ELLMOS_BUNDLE_CONFIG or with the optional bundleConfigPath input on bundle tools. A bundle file is a JSON object with schemaVersion and a bundles array. Each bundle needs id, title, description, and keywords.

This is the basis for future tool-bloat management: instead of exposing many individual tools immediately, an agent can first choose the capability bundle that fits the task.

Tool Catalog

controlcenter_list_tools can start local stdio MCP servers or resolved Claude profile servers and call the standard MCP list_tools request. Profile scans support arbitrary stdio commands, including non-Node launchers, and URL-based remote configs using Streamable HTTP or legacy SSE. The scan is explicit, uses a per-server timeout, does not call any reported tool, and closes each spawned local server after reading the tool list.

controlcenter_build_catalog accepts includeTools: true to persist the same probe results alongside the local server catalog.

controlcenter_assign_tool_bundles compares probed tool names, titles, descriptions, server names, source, and transport metadata with capability-bundle keywords, then reports which tools belong to bundles such as filesystem, software, automation, or control plane.

Profile Audit

controlcenter_audit_profile is the first small policy layer. It currently flags:

• npx starts

• environment variables in server configurations

• missing or invalid server commands

• sensitive name fragments in arguments

Environment values are never printed.

Policy rules are loaded from data/policy-rules.json by default. The file can disable individual rules or override their severity, and controlcenter_audit_profile also accepts a policyConfigPath input for one-off audits.

Project Structure

ellmos-controlcenter-mcp/
|-- src/
|-- test/
|-- data/
|-- README.md
|-- README_de.md
|-- START.md
|-- ARCHITECTURE.md
|-- STATE.md
|-- DECISIONS.md
`-- TODO.md

Documentation

ellmos-ai Ecosystem

This MCP server is part of the ellmos-ai ecosystem — AI infrastructure, MCP servers, and intelligent tools.

MCP Server Family

AI Infrastructure

Desktop Software & Companion Tools

Our partner organization open-bricks bundles AI-native desktop applications — a modern, open-source software suite built for the age of AI.

License

MIT - Lukas Geiger (ellmos-ai)

Bundles and partners

ControlCenter MCP remains a standalone, published MCP server. In the V4 composition it is an optional MCP access surface of the ellmos-core-discovery-bundle: it exposes local MCP-stack, profile, tool and skill discovery to people and MCP-capable clients. It is not the functional owner of policies, decisions, memory, automations, system maps, or the modules behind the discovered tools.

Configured component registries, local MCP servers, profile files and skill libraries are discovery partners, not bundled ownership transfers. The published ControlCenter identity and package name remain unchanged. ControlRoom is a separate planned operator stack, not a rename or a hidden replacement for this server.

Authoritative bundle membership, versions, profiles and any private composition recipes remain in the corresponding bundle manifests. This public section is discovery-only.