ellmos ControlCenter

ellmos ControlCenter MCP

DE Deutsche Version

Part of the ellmos-ai family.

[!NOTE] LLM / AI Agent Integration: This repository provides an llms.txt index file for context optimization, RAG discovery, and agent navigation.

---

Quick Navigation

Installation • Target Personas • Comparative Matrix • System Architecture • Control & Gateway Flow • Governance Invariants • Status • Tools (34) • Gateway • Capability Bundles • Profile Switching • Host Registers • Dashboard • Third-Party Licenses • Documentation • Security Policy • llms.txt Context • Ecosystem Matrix

---

An advanced Model Context Protocol (MCP) administration server and policy-gated gateway 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, provides host-level register mirrors (locks, permissions, resources, governance), and provides an optional local dashboard.

Architecture & Dual Role — Control Plane + Policy-Gated Gateway: ControlCenter combines two complementary operational surfaces:

1. Control Plane (Administration & Configuration): Inventories local servers, resolves profiles, manages capability bundles, mirrors system governance/locks, and generates configurations (controlcenter_switch_profile, controlcenter_build_catalog).
2. Policy-Gated Gateway (On-Demand Tool Invocation): Via controlcenter_invoke and controlcenter_list_available_tools, agents can list and invoke tools on backend MCP servers that the host agent has not loaded into its active context — strictly bounded by pattern-based policy rules (data/gateway-policy.json), argument auditing, and secret scrubbing.

Administration and discovery are local-first with no telemetry or background egress. An explicitly configured gateway call may contact a remote HTTPS backend; redirects are refused and an optional host allowlist can narrow the destination set. Lock and permission tools fail closed when their authority sources cannot be verified.

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.

ControlCenter provides discovery, profile visibility, dashboard workflows, capability bundles, profile-aware tool-list probes, tool-bundle assignments, internationalization, policy audits, host register mirrors (locks, permissions, resources, plans), and the policy-gated gateway (controlcenter_invoke). Hardened for multi-OS deployment with a 48-hour security response SLA.

Target Personas & Discoverability

Persona	Core Profile & Tech Stack	Architectural Friction & Pain Point	How ControlCenter Solves It
AI Infrastructure Engineers & MCP Tooling Architects	Scaling local fleets of MCP servers across multi-agent environments (Claude Code, Codex, Antigravity, Gemini).	Massive agent context window consumption when dozens of MCP servers are loaded simultaneously; config drift across agent profiles.	Dynamic capability bundles (data/capability-bundles.json), hash-consistent profile resolution (controlcenter_resolve_profile), and on-demand tool probes without active memory overhead.
Multi-Agent Runtime Developers & Swarm Operators	Orchestrating autonomous agent loops and workflows (BACH, USMC, LangChain, AutoGen, CrewAI).	Lack of dynamic runtime tool access for tools not pre-declared at agent startup; danger of background process leakage.	Connect-per-call policy-gated gateway (controlcenter_invoke) allowing agents to invoke unloaded backend MCP tools with zero lingering zombie processes.
Enterprise SecOps & Compliance Officers	Auditing local developer environments, sensitive credentials, and agent autonomy boundaries.	Prompt injection attacks via untrusted tool outputs, API key leakage in stack traces, and unmonitored tool executions.	Local-first administration with no telemetry or background egress, explicit HTTPS boundaries for remote gateway calls, fail-closed policy loading (data/gateway-policy.json), bounded credential sanitization, untrusted data wrapping, and append-only audit trails (gateway-audit.jsonl).
Local Homelab Automators & AI Power Users	Managing desktop agents, workflow engines (n8n), and local developer tools.	Fragmented tooling, opaque agent permissions, conflicting locks, and lack of a central visual overview of active MCP stacks.	Centralized local web dashboard (127.0.0.1:3737), bidirectional i18n (EN/DE), and unified host register mirrors for locks (LOCK*.txt), permissions (LOCK.permissions.json), and decisions.

Discovery Keywords & High-Intent Topic Tags: mcp-control-plane, model-context-protocol, mcp-gateway, claude-code-profiles, policy-gated-execution, no-telemetry, local-first-ai, secret-scrubber, multi-agent-coordination, capability-bundles, fail-closed-security, mcp-audit-logging, developer-tools.

---

Comparative Matrix vs Alternatives

Architectural Criterion	ellmos ControlCenter MCP	Static MCP Configurations (claude_desktop_config.json)	Monolithic MCP Meta-Servers	Heavyweight Agent Frameworks (LangChain / CrewAI)	Cloud LLMOps & Remote Gateways
Architecture & Role	Dual Control Plane + Ephemeral Gateway	Static JSON file	Single massive combined process	Embedded code framework	Remote hosted SaaS / proxy
Token & Context Efficiency	Dynamic On-Demand Tool Invocation (controlcenter_invoke)	Poor (All tools must be pre-loaded into context)	Extreme bloat (Dozens of tools in prompt)	Varies (Tools loaded into Python process memory)	Network payload overhead
Process Lifecycle	Connect-Per-Call stdio + optional Windows Job Object (descendant-safe cleanup)	Always-on persistent background daemons	Single monolithic background process	Tied to application execution thread	Cloud-hosted containers
Network Egress & Privacy	Local-first; no telemetry/background egress; explicit remote HTTPS gateway targets are supported	Local stdio / HTTP	Local stdio	Depends on cloud LLM integrations	High egress (Tool data sent to cloud servers)
Policy Gating & Hardening	Fail-Closed Pattern Rules + Recursive Secret Scrubbing	None (Direct unrestricted host access)	Rare / Custom ad-hoc filtering	Inconsistent application-level checks	Organization-level cloud IAM
Untrusted Data Isolation	Enforced GFM Banners for Tool Outputs	None (Raw strings fed directly to LLM)	None	Manual prompt templates	Cloud provider sandboxing
Host Governance Awareness	Native Multi-Agent Locks (LOCK*.txt) & Permissions	None	None	None	None
Profile & Stack Management	Extends Chains, Dynamic Bundles & Catalog Probes	Manual JSON editing	Hardcoded server arrays	Programmatic Python definitions	Web dashboard configuration
Internationalization (i18n)	Bilingual Core (English & German runtime output)	English only	English only	English only	English only
Security SLA & Supply Chain	48h SLA, Zero Transitive Telemetry, Audited Licenses	Vendor dependent	Unaudited third-party tools	Broad attack surface (100+ pip packages)	Third-party vendor trust

---

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

Control Plane & Gateway Lifecycle

sequenceDiagram
    autonumber
    actor Agent as MCP Client (Claude / Codex / Gemini)
    participant CC as ControlCenter MCP Server
    participant Res as Profile & Capability Resolver
    participant Gate as Gateway Policy Guard
    participant Backend as Backend MCP Server (Unloaded)
    participant Scrub as Hardening & Secret Scrubber
    participant Audit as Audit Logger (JSONL)

    Note over Agent,CC: 1. Administration & Profile Discovery
    Agent->>CC: controlcenter_suggest_profile / resolve_profile
    CC->>Res: Inspect ~/.claude/profiles & extends chains
    Res-->>CC: Resolved MCP Configuration & Bundles
    CC-->>Agent: Suggested Profile & --mcp-config flags

    Note over Agent,CC: 2. Policy-Gated Gateway Execution
    Agent->>CC: controlcenter_invoke(server, tool, args)
    CC->>Gate: Evaluate data/gateway-policy.json
    alt Policy Denied or Missing
        Gate-->>CC: Policy Refusal (Fail-Closed)
        CC->>Audit: Log refusal (names only, 0 values)
        CC-->>Agent: Error: Tool / Server denied by policy
    else Policy Allowed
        Gate-->>CC: Dispatch Approved
        CC->>Backend: Connect-per-call (stdio / Streamable HTTP)
        Backend-->>CC: Raw Tool Output / Response
        CC->>Backend: Terminate process / Close transport
        CC->>Scrub: Recursive Secret Redaction & Finite Budgets
        Scrub-->>CC: Sanitized Payload & Truncation Status
        CC->>Audit: Append structured audit event (gateway-audit.jsonl)
        CC-->>Agent: Safe Tool Result wrapped with Data Banners
    end

Governance & Runtime Invariants

ControlCenter enforces 10 architectural and runtime invariants to guarantee local-first administration, explicit outbound boundaries, fail-closed policy loading, and multi-agent coordination resilience across environments:

ID	Invariant	Description	Enforcement Mechanism
INV-LOCAL-01	Local-First & Explicit Egress	Discovery, profile resolution, catalog indexing, and the dashboard execute locally with no telemetry or background egress. Explicit gateway calls may reach remote HTTPS backends.	Dashboard binds to loopback (127.0.0.1:3737); remote HTTP is refused, redirects are refused, and an optional host allowlist narrows HTTPS targets.
INV-GATE-02	Fail-Closed Gateway Policy Guard	Remote/unloaded tool invocations via controlcenter_invoke strictly require pattern authorization in data/gateway-policy.json.	Missing, unreadable, or invalid policy files immediately refuse execution (fail closed).
INV-SUB-03	Ephemeral Child Process Boundaries	Backend stdio processes for probed and invoked MCP servers are spawned on-demand per call and terminated immediately in a finally block.	connect-per-call architecture prevents lingering background zombie processes.
INV-SCRUB-04	Bounded Result Redaction & Finite Budgets	Results are traversed recursively: narrow credential patterns are redacted everywhere, while key-based wiping applies only to structured metadata so requested content is not silently rewritten. Requests are size-bounded and argument values are omitted from the audit log.	Recursive narrow-pattern redaction, structured-metadata key scrubbing, maximum depth and content-block caps, 256 KB request and 1 MB response defaults.
INV-PRIV-05	Non-Elevation / RunAsInvoker	Server operates strictly in unprivileged user space. Never prompts for root/admin elevation.	Operates without root or UAC elevation across Windows, macOS, and Linux.
INV-LOCK-06	Canonical Multi-Agent Lock Awareness	Respects system-wide LOCK*.txt, LOCK.user.*, and LOCK.until.* tokens fail-closed.	Inspects lock trees via host Python lock utilities; unconfigured returns unknown.
INV-PERM-07	Nearest Permission Register Introspection	Evaluates nearest LOCK.permissions.json up directory trees (deny > ask > allow > default).	Hierarchical resolution without granting synthetic permissions or modifying state.
INV-GOV-08	Read-Only Host Governance Federation	Mirrors pending decisions, policies, strategic plans, and BYUM metadata read-only.	Strict scalar projection; never adopts, executes, mutates, or silences items.
INV-SYNC-09	Cloud-Sync Conflict Hardening	Protects repository against multi-host conflict copies and stray lock files.	Comprehensive .gitignore covering *.sync-conflict-*, *-CONFLIT-*, and LOCK.*.
INV-SLA-10	48-Hour Response & 5-Day Triage SLA	Vulnerability reports receive prompt maintainer response and triage commitments.	Documented in SECURITY.md with direct maintainer and umbrella security contacts.

Status

• Phase: Alpha
• Version: 0.7.4
• 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, early audits, and read-only host governance metadata

Tools

Tool	Purpose
controlcenter_status	Show stack, profile, and detected-server status
controlcenter_actual_self_receipt	Run a native self list_tools probe and emit a short-lived signed runtime receipt when explicitly configured
controlcenter_get_language	Show the current ControlCenter output language
controlcenter_set_language	Set the ControlCenter output language for this running server instance
controlcenter_list_local_servers	Scan local MCP repositories below the MCP root and enrich them with kind and state ownership from mcps.catalog.v1.json
controlcenter_describe_mcp	Describe one MCP server from mcps.catalog.v1.json: kind, namespace, state ownership, wrapping, and composition
controlcenter_list_stacks	Read registered stacks from stacks.catalog.json and validate their ellmos.stack.v2 manifests
controlcenter_describe_stack	Describe typed components, roles, policies, and validation warnings for one registered stack
controlcenter_context_pack	Build a bounded, manifest-only handoff for a registered stack at short, execution, or full detail
controlcenter_list_tools	Start local or profile-defined MCP servers and read their real list_tools output
controlcenter_find_capability	Rank typed native-binding claims from a hash-consistent System Explorer resolution without selecting or executing one
controlcenter_tool_overview	Show resolution-bound component claims while keeping declared and runtime-state axes separate
controlcenter_assign_tool_bundles	Assign probed MCP tools to capability bundles
controlcenter_list_bundles	Group local servers by capability bundle
controlcenter_suggest_bundles	Recommend bundles for a task
controlcenter_list_profiles	List MCP profiles from the profile root (defaults to ~/.claude/profiles; override with ELLMOS_PROFILE_ROOT)
controlcenter_suggest_profile	Recommend a profile for a task
controlcenter_resolve_profile	Resolve a profile including extends chains
controlcenter_switch_profile	Prepare a generated --mcp-config file and configurable launch command
controlcenter_audit_profile	Run initial policy checks against a profile
controlcenter_build_catalog	Build a JSON catalog of local MCP servers, optionally including tool probes
controlcenter_list_skills	Inventory deployed skills (~/.claude/skills by default; Claude Code convention, override with ELLMOS_SKILLS_ROOT) and the source skills library
controlcenter_find_skill	Match keywords for a task or intent against the scanned skill catalogue and return ranked candidates — see Querying skill search
controlcenter_resolve_semantic_route	Validate an LLM/user-selected role, expert and persona against a provider-neutral map and verify endpoints against the live skill inventory
controlcenter_list_plugins	Inventory installed plugins (~/.claude/plugins by default; Claude Code convention, override with ELLMOS_PLUGINS_ROOT) and local ellmos modules
controlcenter_list_locks	List active LOCK*.txt project locks across the configured roots — see Host registers
controlcenter_check_lock	Check whether one path is locked, including locks inherited from parent directories
controlcenter_evaluate_permission	Report what the nearest LOCK.permissions register allows an agent to do at a path
controlcenter_list_decisions	List pending user decisions by identifier, date, title and status
controlcenter_list_governance	Federate allowlisted decision, policy, strategic-plan and BYUM metadata read-only; report each source separately and never adopt or execute a candidate
controlcenter_list_resources	List rows from the host's resource inventory (systems and/or installed software) — read-only mirror; the register's authority sits with the ControlRoom programme, not here
controlcenter_describe_resource	Full row detail for one resource by its inventory id, from the same read-only mirror
controlcenter_list_available_tools	List the tools of MCP servers this host has not loaded, without loading them — see Gateway
controlcenter_invoke	Run one tool on a server this host has not loaded and return its result, policy-gated and audited

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. On Windows, configure ELLMOS_PROCESS_SUPERVISOR with the local Job-Object supervisor to bind descendants as well as the direct stdio child. The cost is roughly 200–500 ms per call on a cold stdio server; the result is bounded cleanup instead of an unverified direct-child-only kill.

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

Outcome	Meaning
unknown-server	The name is not in the addressable set. The known names are returned.
unreachable	The server exists but could not be asked. Not "returned nothing".
unknown-tool	The server has no such tool. Its available tool names are returned, so a wrong guess self-corrects in one step.
target-error	The call arrived and the target reported a tool error. This is a backend result, not a ControlCenter failure.

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:

Control	Behaviour
Recursive redaction	Narrow credential shapes (sk-, ghp_, AKIA, JWT, …) are replaced everywhere, at every nesting level. Secret-named keys (auth, token, apiKey, …) are wiped whole only in structuredContent — never in content blocks, which carry the payload the caller asked to read. The result reports how many values changed. Disable with redactResults: false — a deliberate weakening.
Request budget	Oversized arguments are refused, never shortened; a truncated argument set would silently change the request. ELLMOS_GATEWAY_MAX_REQUEST_BYTES, default 256 KiB.
Response budget	Oversized answers are truncated and flagged, so the part that arrived stays usable. ELLMOS_GATEWAY_MAX_RESPONSE_BYTES, default 1 MiB.
Nesting and blocks	ELLMOS_GATEWAY_MAX_DEPTH (32) and ELLMOS_GATEWAY_MAX_CONTENT_BLOCKS (200). Cycle-safe, so a self-referential payload cuts off instead of looping.
Concurrency	ELLMOS_GATEWAY_MAX_CONCURRENT (4). Without it a parallel batch would spawn one backend process each. A call that gets no slot is refused, not queued forever.
Transport	HTTPS only; plain HTTP allowed on loopback alone. Redirects refused. Narrow further with allowedRemoteHosts (supports *. subdomains).
Untrusted marking	Forwarded content is fenced with a banner marking it as data, not instructions — the gateway pipes third-party output into an agent's context.

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.

Catalog	Schema	Root (env override)	Used by
modules.catalog.json	ellmos.modules-catalog.v1	.AI/.MODULES (ELLMOS_MODULES_ROOT)	controlcenter_list_plugins
stacks.catalog.json	ellmos.stacks.catalog.v1	.AI/.STACKS (ELLMOS_STACKS_ROOT)	controlcenter_list_stacks, controlcenter_describe_stack, controlcenter_context_pack
mcps.catalog.v1.json	ellmos.mcps.v1	.AI/.MCP (ELLMOS_MCP_CATALOG)	controlcenter_list_local_servers, controlcenter_describe_mcp, controlcenter_status

The MCP catalog contributes what a directory scan cannot see: mcp_kind (tool, adapter, stack, control-plane), whether a server keeps persistent state, which component owns that state per namespace, and optional declared capability tags. A tag block is versioned as capability_tags: {"schema":"ellmos.capability-tags.v1","tags":["catalog","read-only"]}; tags are normalized to lower case, sorted, and treated as metadata only. Duplicate, mistyped, or malformed tags make the catalog explicitly invalid instead of silently dropping data. 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, foreign-schema, or structurally invalid 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, governance, resources

The seven 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 six host-local registers — project locks, an agent-neutral permission register, a pending decision list, a policy registry, the strategic-plan index, and a resource inventory of systems and installed software.

controlcenter_list_governance composes the generated decision index, the existing ellmos.plans-register/1 strategic-plan index, and an explicitly configured ellmos.policy-registry.v1 file. The policy registry is validated only through the canonical PolicyRegistry.load() API. Every source reports available, unconfigured, unreadable, or invalid; partial data never claims completeness, and a valid registry with zero BYUM candidates reports an honest zero. Plan paths, notes and host variants remain in _PLANS; BYUM rows remain pending advisory pointers without adoption or execution authority.

controlcenter_list_resources and controlcenter_describe_resource are a read-only mirror of .SYNC/_inventory/inventory.db. Register authority sits with the ControlRoom programme's own resources.inventory resolver role, not with this server — this mirror can go stale between syncs and never claims otherwise.

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:

Variable	Purpose
ELLMOS_LOCK_SCRIPTS	Directory holding the canonical lock_utils.py, permissions.py and lock_scan.py. Required by the three lock and permission tools.
ELLMOS_LOCK_ROOTS	Optional path to lock_roots.json. Defaults to the file beside the lock scripts.
ELLMOS_DECISIONS_ROOT	Directory holding the decision chain and its generated index. Required by controlcenter_list_decisions.
ELLMOS_INVENTORY_DB	Path to the resource inventory SQLite file (.SYNC/_inventory/inventory.db). Required by controlcenter_list_resources and controlcenter_describe_resource.
ELLMOS_POLICY_REGISTRY_PATH	Explicit path to an ellmos.policy-registry.v1 registry. Required for the policy side of controlcenter_list_governance.
ELLMOS_POLICY_REGISTRY_SRC	Optional source root containing the canonical policy_registry Python package.
ELLMOS_PLANS_REGISTER	Explicit path to _control-center/_PLANS/plans-register.json in schema ellmos.plans-register/1. Required for the plan side of controlcenter_list_governance.
ELLMOS_PYTHON	Interpreter to run the bridge with. Defaults to python, falling back to python3.

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.

controlcenter_list_governance uses fixed field allowlists. It never returns source URIs or plan paths, host variants, plan notes, questions, options, recommendations, rationale, prompts, full text, reasons, secure/avatar content, action payloads, execution payloads, or receipts, and it never dereferences a registry pointer.

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.

	Query	Top result
❌	My program crashes when saving and I don't know why	mcp-config-sync (score 6 — matched on when, know, why)
✅	debug bug test failure	bugfix-protocol (score 5 — matched on bug, debug)

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 versioned header/auth contract is ellmos.tool-scan-headers.v1: configured static headers are passed to the SSE event-stream GET and every MCP message POST; the installed SDK transport contract is used, redirects are refused, and configured header/environment values are masked from probe errors. Every transport is closed after the probe.

The scan is explicit, uses a per-server timeout, does not call any reported tool, and applies finite defaults of four concurrent probes and a 1 MiB aggregate response budget. maxParallelProbes is bounded to 1–32 and maxResponseBytes to 1 KiB–16 MiB. A budget-exhausted or not-started target is returned as status: incomplete with an unknown tool count; it is never reported as a successful empty server. These controls are available on controlcenter_list_tools, controlcenter_assign_tool_bundles, controlcenter_build_catalog, and the dashboard scan.

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

Third-Party Licenses & Transparency

This project adheres strictly to 100% permissive open-source licensing across all direct runtime and development dependencies:

• 0% Copyleft / GPL / AGPL exposure.
• Fully audited and compatible with commercial, enterprise, and local-first deployments.
• Audited direct dependencies: @modelcontextprotocol/sdk (MIT), zod (MIT), typescript (Apache-2.0), vite/vitest (MIT), @types/node (MIT), and @emnapi/core/@emnapi/runtime (MIT).

Full SPDX license texts, copyright notices, and compliance attestations are documented in THIRD_PARTY_LICENSES.md. Canonical copyright and ecosystem attribution is declared in NOTICE.

Documentation

For...	Read...
Quick start	START.md
Current state	STATE.md
Architecture	ARCHITECTURE.md
Roadmap	ROADMAP.md
Decisions	DECISIONS.md
Open tasks	TODO.md
Changes	CHANGELOG.md
Notice & attribution	NOTICE
Third-party licenses	THIRD_PARTY_LICENSES.md
Marketing & discoverability	MARKETING-LOG.txt
LLM crawler summary	llms.txt

Security Policy & Vulnerability Reporting

ControlCenter adheres to a strict multi-agent security model. See SECURITY.md for full details on:

• Zero-Egress & Local-First Guarantees: Local execution with no telemetry.
• Fail-Closed Gateway Policy: Pattern-based enforcement and bounded secret scrubbing.
• 48-Hour Response SLA: Binding response commitment (INV-SLA-10) via security@open-bricks.org and security@ellmos.ai.

llms.txt Context Index

For automated LLM agent integration, RAG crawling, and prompt optimization, ControlCenter provides a structured llms.txt index file at the repository root. It summarizes tool schemas, governance invariants, CLI usage patterns, and ecosystem relationships in an LLM-friendly format.

ellmos-ai Ecosystem, Statutory Liability (§ 521 BGB) & Security SLA

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

MCP Server Family

Server	Tools	Focus	npm
FileCommander	47	Filesystem, process management, interactive sessions, cloud-lock-safe operations	ellmos-filecommander-mcp
CodeCommander	22	Code analysis, JSON repair, imports, diffs, regex	[ellmos-codecommander-mcp](https://www.npmjs.com/pac