> ## Documentation Index
> Fetch the complete documentation index at: https://docs.unclerobertconsulting.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Ops Agent settings and MCP integrations

> Tune Ops Agent LLM hyper-parameters, manage built-in OS integrations, register custom webhooks, and mount Model Context Protocol (MCP) servers.

## Overview

The Settings page gives you fine-grained control over how the Ops Agent thinks and what it can connect to. Two sections matter most:

* **Ops Agent LLM Controls** — persona, model selection, sampling hyper-parameters, and cognitive governance toggles.
* **Integrations & MCP** — built-in OS integrations, custom webhooks, and Model Context Protocol (MCP) servers, each with a live connection test.

Use this page when you need to:

* Make the Ops Agent more deterministic or more creative for a given workload.
* Set a fallback model so runs survive a primary model outage.
* Register an external webhook or mount an MCP server as an agent tool source.
* Verify that an integration handshake works before relying on it in a workflow.

## Ops Agent LLM controls

### Identity and tone

Define who the agent is and how it communicates:

| Control                           | What it does                                                                                                                                                     |
| :-------------------------------- | :--------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Orchestrator Name**             | Display name for the orchestrating agent.                                                                                                                        |
| **Persona & Executive Role**      | Free-text role the agent adopts, for example `Chief Systems Architect & Autonomous Swarm Director`.                                                              |
| **Tone & Operating Code**         | One of four styles: Servant Leadership, Technical Architect, Executive Operator, or Concise Action-Only.                                                         |
| **Active Domain Knowledge Brain** | Knowledge base the agent reasons over, such as the URC 7-Department Canonical Brain or the [Founder Signal](/workflows/founder-signal-system) accelerator brain. |

### Model and sampling

Control which model runs and how it generates output:

| Control                      | What it does                                                                                                                                          | Default          |
| :--------------------------- | :---------------------------------------------------------------------------------------------------------------------------------------------------- | :--------------- |
| **Primary Reasoning Model**  | Main model for orchestration. Options include Gemini 2.5 Flash, Gemini 2.5 Pro, GPT-4o, Claude 3.5 Sonnet, and Llama 3.3 70B.                         | Gemini 2.5 Flash |
| **Automatic Fallback Model** | Model used if the primary fails. Choose `None` for strict fail-fast behavior.                                                                         | `gemini-2.5-pro` |
| **Sampling Temperature**     | Slider from 0.0 to 1.0. Below 0.3 is deterministic strict-SOP adherence, 0.3–0.7 balances strategy and execution, above 0.7 favors creative ideation. | `0.2`            |
| **Max Output Tokens**        | Upper bound on tokens per generation.                                                                                                                 | `4096`           |
| **Top-P Nucleus Sampling**   | Probability mass cutoff for token sampling.                                                                                                           | `0.9`            |

### Cognitive governance

Three toggles gate how the agent reasons. All three are on by default:

* **Anti-Passivity Mandate (Zero Refusal Gate)** — the agent must attempt the task rather than deflect or defer.
* **Multi-Step Chain-of-Thought DAG Planning** — the agent decomposes work into a multi-step plan before executing.
* **Quality Flywheel Auto-Evaluation & Scoring** — the agent scores its own output and feeds the score back into future runs.

Click **Save** to persist the configuration. Saving updates the orchestrator name, default model, and system prompt on the server.

## Secrets vault

The Secrets Vault (API Keys) tab stores provider credentials and bridges them to the live environment. When you save a secret, the platform applies it to the running process immediately. No restart is required, and a connection test run right after saving uses the new credential.

To add a secret, enter a provider name and the secret value, then click **Save**. The vault:

* Normalizes the provider name to its canonical environment key. For example, `hubspot`, `hubspot_pat`, and `hubspot_access_token` all map to `HUBSPOT_PAT`; unrecognized providers map to the uppercased provider name.
* Stores only a masked preview of the value, in the form `sk_1••••••••e0a1`. Values of 8 characters or fewer show as a fully masked placeholder.
* Increments the secret's version each time you update an existing provider.

Recognized provider mappings:

| Provider name contains  | Environment key                |
| :---------------------- | :----------------------------- |
| `instantly`             | `INSTANTLY_API_KEY`            |
| `hubspot`               | `HUBSPOT_PAT`                  |
| `elevenlabs`            | `ELEVENLABS_API_KEY`           |
| `agentmail`             | `AGENTMAIL_API_KEY`            |
| `google` or `gemini`    | `GOOGLE_GENERATIVE_AI_API_KEY` |
| `openai`                | `OPENAI_API_KEY`               |
| `anthropic` or `claude` | `ANTHROPIC_API_KEY`            |
| `stripe`                | `STRIPE_SECRET_KEY`            |

### Automatic provider sync

The vault also syncs in the other direction. Whenever you open the secrets or integrations list, the platform scans the environment for five core provider keys and registers any it finds as vault entries and integrations: Instantly (`INSTANTLY_API_KEY`), HubSpot (`HUBSPOT_PAT`), ElevenLabs (`ELEVENLABS_API_KEY`), AgentMail (`AGENTMAIL_API_KEY`), and Google Gemini (`GOOGLE_GENERATIVE_AI_API_KEY`). Credentials configured outside the app show up in the vault with a masked preview and a `connected` status without any manual registration.

## Integrations & MCP hub

The Integrations & MCP tab manages three kinds of connections: built-in OS integrations, custom webhooks, and MCP servers.

### Built-in OS integrations

Eight integrations ship pre-configured:

| Integration                          | Purpose                                                                  | Protocol                     |
| :----------------------------------- | :----------------------------------------------------------------------- | :--------------------------- |
| **HubSpot CRM & Sales**              | Two-way lead capture, deal pipeline sync, contact engagement tracking.   | OAuth 2.0 / REST API         |
| **AgentMail Inbound & Direct SMTP**  | Autonomous cold outreach and verified lead reply ingestion.              | Direct REST / Webhook Ingest |
| **ElevenLabs / Pamela Telephony**    | Inbound triage, diagnostic slot booking, post-call transcript ingestion. | REST & Webhook Handshake     |
| **IONOS Cloud & Domains**            | Enterprise DNS, hosting, sovereign instances, and SSL certificates.      | Certified Partner Network    |
| **Microsoft 365 Operating Backbone** | OneDrive ledger sync, Outlook calendar dispatch, M365 reconciliation.    | M365 Graph Bridge            |
| **Twilio Telephony & SIP**           | Local number routing, SMS gateway, SIP forwarding.                       | Twilio REST API v2010        |
| **n8n Autonomous Workflow Engine**   | Scheduled CRON triggers, sync loops, background jobs.                    | n8n Webhook / MCP Server     |
| **Pulse Social Scheduler**           | Autonomous content queue scheduling and distribution.                    | AgentLab Social Dispatch API |

### Add a custom webhook

Click **Add Webhook** and provide a name, an endpoint URL, and an optional API key. The integration is stored with type `webhook`:

```json theme={null}
{
  "name": "Mercury Bank Webhook",
  "type": "webhook",
  "config": {
    "endpoint": "https://api.mercury.com/v1/webhooks",
    "apiKey": "sec_test_123"
  },
  "status": "active"
}
```

### Mount an MCP server

Click **Add MCP Server** to mount a standard-protocol tool source such as PostgreSQL MCP, BigQuery MCP, or Filesystem MCP. Configure:

* **Server Identifier Name** — a label like `Postgres-MCP`.
* **Transport Protocol** — `sse` (Server-Sent Events URL) or `stdio` (local command).
* **Endpoint or command** — an SSE URL such as `https://mcp.agent-lab.tech/sse`, or a CLI command such as `npx -y @modelcontextprotocol/server-postgres` for stdio.
* **Capabilities** — comma-separated list, for example `tools,resources`.
* **Authorization Token / API Key** — optional.

The integration is stored with type `mcp`:

```json theme={null}
{
  "name": "PostgreSQL MCP",
  "type": "mcp",
  "config": {
    "transport": "sse",
    "endpoint": "https://mcp.agent-lab.tech/sse",
    "capabilities": ["tools", "resources"]
  },
  "status": "active"
}
```

### Integration status

Every integration carries a `status` field. New integrations default to `active` unless you supply a different status when creating or updating them.

## Test a connection

Each registered webhook and MCP server has a **Test** action that runs a live handshake through the `settings.testIntegration` mutation. The test picks the deepest check available for the integration:

1. **Provider-verified checks.** Integrations whose name matches a known provider are validated against that provider's live API using the credentials configured in the environment or [secrets vault](#secrets-vault):

   * **Instantly** verifies the outbound connection over `REST API v2 (Instantly Outbound)` using `INSTANTLY_API_KEY`.
   * **ElevenLabs** (names containing `elevenlabs` or `pamela`) calls the ElevenLabs user endpoint with `ELEVENLABS_API_KEY`.
   * **HubSpot** queries the CRM contacts endpoint with `HUBSPOT_PAT` (also accepting `HUBSPOT_ACCESS_TOKEN`, `HUBSPOT_DEVELOPER_API_KEY`, or `HUBSPOT_API_KEY`).

   If the required key is missing, the test fails with a message naming the key, for example `ELEVENLABS_API_KEY is not configured in environment or vault.`
2. **Live HTTP ping.** If the integration config includes an `http://` or `https://` endpoint URL, the test sends a `HEAD` request with a 5-second timeout and reports the actual HTTP status and measured roundtrip latency.
3. **Local fallback.** Integrations with no reachable URL are verified as a local profile and report `Local Plugin / Runtime Hook` (or `Model Context Protocol v1.0 (stdio)` for MCP).

A successful live ping returns the protocol negotiated, the HTTP status, and the roundtrip latency:

```json theme={null}
{
  "success": true,
  "latencyMs": 34,
  "protocol": "Model Context Protocol v1.0 (SSE/stdio)",
  "message": "HTTP 200 OK from https://mcp.agent-lab.tech/sse (34ms roundtrip).",
  "timestamp": "2026-09-07T20:06:13.000Z"
}
```

A failed handshake returns `success: false` with the underlying error, for example an HTTP error code or a timeout message, so you can tell a bad credential apart from an unreachable endpoint.

Run a test after adding an integration, and again whenever a workflow that depends on it starts failing, to confirm the endpoint still responds.