---
title: "AI assistant, MCP tokens and coding-agent keys"
description: "Open the Backoffice assistant, set model routing and provider keys, mint MCP tokens with read scopes, and create coding-agent keys, with the limits Backoffice shows."
canonical_url: "https://noxtica.com/docs/backoffice/ai-assistant-and-mcp/"
markdown_url: "https://noxtica.com/docs/backoffice/ai-assistant-and-mcp/index.md"
last_updated: 2026-10-27
published: 2026-10-27
section: "backoffice"
section_title: "Backoffice Guide"
site: "Noxtica"
language: "en"
source: markdown
---

# AI assistant, MCP tokens and coding-agent keys

Backoffice has three AI-related surfaces. The **assistant** is a chat panel for operators. **MCP Tokens** let an outside AI client read workspace data through the Model Context Protocol. **Coding Agent Keys** let a coding agent read your errors, traces and logs. A workspace Admin sets all three up and keeps each one to the narrowest access it needs.

## What it is

Availability: Availability depends on your plan, role and the modules enabled for your organisation.

| Surface | Where | Who uses it | What it reads or does |
| --- | --- | --- | --- |
| Assistant | Floating **Assistant** button on every page | Admin and above | Answers questions about your workspace inside your own role and workspace |
| LLM Routing, Speech, LLM Provider Keys | `/ai/llm-routing`, `/ai/speech`, `/ai/provider-keys` | Admin | Where model calls go, voice features, your own provider keys |
| MCP Tokens | `/integrations/mcp-tokens` | Admin | Bearer tokens for an MCP client |
| Coding Agent Keys | `/integrations/coding-agent-keys` | Admin | Read-only keys for a coding agent |

How they differ from the concept pages: [AI Assistant](https://noxtica.com/docs/concepts/ai-assistant/) describes the operator copilot and [MCP Integration](https://noxtica.com/docs/concepts/mcp-integration/) describes the external read surface. The wire protocol, endpoint and tool list for MCP are in [MCP server](https://noxtica.com/docs/api/mcp/). This page covers the Backoffice side.

Treat every answer as a signal. The assistant says "AI-generated — may be inaccurate or incomplete. Not advice. Verify independently before acting." An MCP or coding-agent response is data for your own tooling to interpret.

## Prerequisites

- The Admin role. Lower roles do not see the assistant button or these pages. See [Backoffice guide](https://noxtica.com/docs/backoffice/overview/).
- An authenticator app. Minting a token or key, and changing assistant configuration, ask for a fresh MFA check.
- The assistant enabled for your organisation. When it is not, opening it shows "Assistant not enabled". When it is enabled but no provider is configured, it shows "Configuration pending". A tenant-level switch can also turn AI off for a workspace.
- For MCP Tokens: MCP enabled for your organisation. The page says the MCP server stays disabled until an operator turns it on.
- A secret store. Tokens and keys are shown once.

## Steps

### 1. Use the assistant

1. Select the **Assistant** button (bottom-right by default). Drag it to another corner, or focus it and press an arrow key to move it.
2. Ask a bounded question about your workspace, for example about policies, domains, fingerprints or the audit log.
3. Use the toolbar for **New chat**, **Save**, **Sessions**, the **Assistant model** picker, **Dock to the right** or **Dock to the bottom**, **Minimize** and **Maximize**. The panel is a helper: it does not block the page behind it.

What to expect:

- The assistant answers within your role and the active workspace.
- Answers show the data they were based on under **Sources**, or "No tenant data was read for this answer."
- Credits: the panel shows your monthly AI credits remaining. When they run out it says so and they reset at the start of the next month (UTC).
- Sessions: **Usage & export** lists your sessions with model, token cost and status, never their contents. When available, conversation text is kept for 30 days in a size-capped record that only you can see; you can export or erase it. If retention is off for your workspace, text is never stored on the server.
- Voice: when speech is enabled, dictation inserts a transcript for review and sends nothing until you press **Send**. Read-aloud plays an answer.
- Changes: write actions are off by default. Where they are enabled, the assistant only proposes a change. A **plan** shows the diff, and a person approves it with a fresh MFA check. The assistant cannot apply a change itself.

### 2. Choose model routing (LLM Routing)

`AI & Agents → LLM Routing` offers three modes: **Platform default**, **Prefer local** (use a locally running model when one is available, otherwise fall back silently) and **Require local (data residency)** (never send the workspace's data to a cloud model; a task with no matching local model fails). The section describes the setting as the model choice for the workstation Backoffice Agent's tasks for this workspace. Pick **Save**.

### 3. Set speech and your own provider keys

- `Speech`: shows which voice features are available and why any are off. Add your own speech provider key if you do not want to use the platform pool. Usage appears on **Usage & Quotas**.
- `LLM Provider Keys`: add a key for a provider you have an account with. Keys are stored encrypted and never shown again. Choose a **Data policy** for each key: **Tier-safe** (eligible to serve assistant turns once it has a model) or **Dev only** (can be tested, cannot serve assistant turns; OpenRouter is always dev only). Use **Discover models**, pick a **Default model**, and **Test** the credential. If you remove a key while using your own keys only, assistant turns stop when no other eligible key can serve them; the platform does not take over.

### 4. Mint an MCP token

1. Open `Integrations & API → MCP Tokens`. Tabs: **Manage** and **Integration** (an in-app guide).
2. Enter a **Name** of 2 to 200 characters that identifies the client that will hold it.
3. Choose read scopes. Scopes are grouped by area: Workspace & usage, Detection analytics, Devices, Noxtica Pulse, Rules & disputes, Alerts, Session replay, Governance, Agent Worker, Published content, RUM, APM & runtime tracker, KYC, KYB & KYM, SEO, Billing & wallet, Workspace settings & integrations, Infrastructure, Team & credential inventory. Each read scope names what it exposes, for example **Device scores & related devices** or **Audit log**.
4. Optionally restrict the tools the token can call, and set an expiry.
5. Select the create button, enter your MFA code, and copy the token. It is shown once.

Rules Backoffice enforces:

- The token is bound to your workspace, and it can never carry a permission that you do not hold.
- Expiry is mandatory and bounded. If you give none, the token expires 90 days after creation. A longer date is reduced to 90 days, and a date in the past is refused. A token that sits idle can also expire.
- A token has a summary or detail tier that changes how fine-grained the read results are. The tier never widens the token's column allowlist.
- Revoke a token to cut off any client at once. The row shows **Last used**.
- Write tokens exist only when write tokens are enabled for your organisation; otherwise minting one is refused. An agent that holds a write token can only plan a change. The plan appears under **Pending AI-agent plans** for 24 hours. You approve it with a fresh MFA check within 10 minutes, or reject it. If the token is revoked, the plan is cancelled.

Connect your client with the endpoint, headers and rate limits in [MCP server](https://noxtica.com/docs/api/mcp/). **Agentic Security → MCP token usage** shows queries used against the period limit. See [Risk actions, investigations and entity lookup](https://noxtica.com/docs/backoffice/risk-actions-and-investigations/).

### 5. Create a coding-agent key

1. Open `AI & Agents → Coding Agent Keys`.
2. Enter a **Label** (for example `ci-agent`; up to 200 characters) and choose scopes. The available scopes are **RUM errors**, **APM traces**, **APM errors**, **APM metrics**, **Logs**, **Profiles**, **Fingerprint / bot signals** and **KYC risk summaries**. Choose only the data the agent needs. The KYC scope covers risk summaries, not identity documents.
3. Select **Mint key** and enter your MFA code. Copy the `cak_` key; it is shown once.
4. Use **Rotate key** to replace the secret with the same scopes and label, or **Revoke key** to end access at once.

A coding-agent key is tenant-wide and read-only. The page says it never grants write access.

## Verify it works

- The assistant answers a question and lists its sources.
- The MCP token appears under **Manage** with scopes and an expiry, and **Last used** changes after your client's first call.
- Your MCP client lists tools and a call returns data inside its scopes. A call outside the scopes is refused.
- The coding-agent key row shows **Created** and, after use, **Last used**.

## Troubleshooting

| Symptom | Cause | Fix |
| --- | --- | --- |
| No **Assistant** button | Role below Admin | Ask a workspace Admin |
| "Assistant not enabled" | Not enabled for your environment | Ask your account team |
| "Configuration pending" | No AI provider key is configured | Ask your account team, or add a provider key if you use your own |
| "You've reached the assistant's usage limit" | A usage or credit limit was reached | Wait for the reset, or ask your account team |
| "Speech input is not available for your account yet" | Speech is off or not granted | See `AI & Agents → Speech` |
| Token creation says to enter a name or choose a scope | Name under 2 characters or no scope selected | Complete the form |
| Token creation is refused for a scope | You do not hold that permission | Ask for the role that includes it |
| A plan expired | Not approved within 10 minutes | Ask the agent to plan again |
| MCP client is refused | The token is revoked, expired or out of scope | Mint a new token with the scopes the client needs |

## Related

- [AI Assistant](https://noxtica.com/docs/concepts/ai-assistant/): what the operator copilot is for.
- [MCP Integration](https://noxtica.com/docs/concepts/mcp-integration/): the external read surface.
- [MCP server](https://noxtica.com/docs/api/mcp/): endpoint, protocol and limits.
- [Integrations and API pages](https://noxtica.com/docs/backoffice/integrations/): the other credential pages.
- [Roles and permissions](https://noxtica.com/docs/security/roles-and-permissions/): the capability list.
- [Know Your Agent](https://noxtica.com/docs/concepts/know-your-agent/): incoming agents, a separate topic.