MCP Tools

Nova supports the Model Context Protocol (MCP) for extending agent capabilities with external tool servers. MCP servers run as subprocesses managed by the Orchestrator, and their tools become available to all agents automatically.

What is MCP?

The Model Context Protocol is an open standard for connecting AI models to external tools and data sources. An MCP server exposes a set of tools (functions) that agents can call during their execution. Nova’s Orchestrator connects to MCP servers at startup and makes their tools available alongside the built-in tools (file I/O, shell, git).

How agents use MCP tools

The Orchestrator loads MCP server configurations from the database at startup
It connects to each enabled server and discovers its available tools
When an agent runs, its tool list carries a one-line capability index entry per connected server (name, description, tool count) instead of every tool’s full schema
When a task involves an integration, the agent calls load_integration_tools(server=...) — the server’s real tool schemas are spliced into the conversation for the rest of that task
The agent can then call any loaded MCP tool by name (prefixed with mcp__<server>__<tool>)
The Orchestrator dispatches the call to the appropriate MCP server and returns the result

Lazy loading and “Always inject”

Tool schemas are heavy: a single integration can expose 50+ tools, and injecting every schema into every LLM call costs thousands of prompt tokens and measurably degrades tool selection on small local models. Lazy loading is therefore the default: an installed integration costs ~15 tokens per call until an agent actually needs it.

If a specific integration is hot enough that the extra load step hurts, expand its server card on the Integrations page and switch on Always inject tools — that server’s schemas then ride along in every LLM call, and it drops out of the capability index. Agents on a pod whose tool allowlist pins specific mcp__ tools also get those schemas injected directly, without a load call.

Available MCP servers

Nova ships with a built-in catalog of pre-configured MCP servers. You can add any of these from the Dashboard’s MCP page with one click.

Core tools

Server	Description	Package
Filesystem	Read, write, and navigate files	`@modelcontextprotocol/server-filesystem`
Git	Inspect and operate on Git repositories — log, diff, status, commit	`mcp-server-git` (uvx)
Memory	Persistent key-value memory store that survives across sessions	`@modelcontextprotocol/server-memory`

Smart home and automation

Server	Description	Package
Home Assistant	Read sensor state and control lights, climate, covers, and locks	`hass-mcp` (uvx)
n8n	List, inspect, create, and run n8n workflows	`n8n-mcp`

Network

Server	Description	Package
AdGuard Home	Query DNS stats and manage filtering, clients, and blocked services	`@samik081/mcp-adguard-home`

Development

Server	Description	Package
GitHub	Manage repos, issues, pull requests, and code search	`@modelcontextprotocol/server-github`
GitLab	Interact with GitLab projects, merge requests, and issues	`@modelcontextprotocol/server-gitlab`
Docker	List, inspect, and manage containers and images on the host	`docker-mcp` (uvx)

Web and search

Server	Description	Package
Brave Search	Web and local search via the Brave Search API	`@modelcontextprotocol/server-brave-search`
Fetch	Fetch URLs and convert web pages to Markdown	`mcp-server-fetch` (uvx)
Firecrawl	Web scraping, crawling, and search with JS rendering	`firecrawl-mcp`
Puppeteer	Browser automation — screenshot, click, fill forms, scrape	`@modelcontextprotocol/server-puppeteer`

Databases

Server	Description	Package
PostgreSQL	Query a PostgreSQL database	`@modelcontextprotocol/server-postgres`
SQLite	Read and query SQLite database files	`@modelcontextprotocol/server-sqlite`

AI and reasoning

Server	Description	Package
Sequential Thinking	Structured multi-step reasoning for complex problem decomposition	`@modelcontextprotocol/server-sequential-thinking`

Communication

Server	Description	Package
Slack	Read channels, send messages, and search Slack workspaces	`@modelcontextprotocol/server-slack`

Infrastructure

Server	Description	Package
Cloudflare	Manage Workers, KV, R2, D1, DNS, and Tunnels	`@cloudflare/mcp-server-cloudflare`
Tailscale	Manage devices, ACLs, DNS, and network configuration	`@hexsleeves/tailscale-mcp-server`

Adding MCP servers via the Dashboard

Navigate to MCP in the Dashboard sidebar
Click Add from Catalog to browse available servers
Select a server and fill in any required configuration (API keys, paths)
Click Add — the Orchestrator will connect to the server and discover its tools
The server immediately appears in every agent’s capability index; agents load its tools on demand (or enable Always inject tools on the server card to skip the load step)

Adding custom MCP servers

You can register any MCP server, not just those in the catalog. From the Dashboard’s MCP page:

Click Add Custom Server
Provide the server name, command, arguments, and any environment variables
The Orchestrator will attempt to connect and discover tools

Or via the API:

curl -X POST http://localhost:8000/api/v1/mcp-servers \
  -H "X-Admin-Secret: your-admin-secret" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "my-server",
    "command": "npx",
    "args": ["-y", "@my-org/mcp-server-custom"],
    "env": {"API_KEY": "..."},
    "enabled": true
  }'

Managing MCP servers

Action	API	Dashboard
List servers	`GET /api/v1/mcp-servers`	MCP page
Add server	`POST /api/v1/mcp-servers`	Add from Catalog / Add Custom
Update config	`PATCH /api/v1/mcp-servers/{id}`	Edit button
Remove server	`DELETE /api/v1/mcp-servers/{id}`	Delete button
Reconnect	`POST /api/v1/mcp-servers/{id}/reload`	Reload button

Environment variables

Some MCP servers require API keys or configuration via environment variables. These are set per-server and passed to the subprocess when it starts. Required variables for catalog entries are documented in the catalog and prompted during setup.

Server	Required variables
AdGuard Home	`ADGUARD_URL`, `ADGUARD_USERNAME`, `ADGUARD_PASSWORD`
Brave Search	`BRAVE_API_KEY`
GitHub	`GITHUB_PERSONAL_ACCESS_TOKEN`
GitLab	`GITLAB_PERSONAL_ACCESS_TOKEN`, optionally `GITLAB_API_URL`
Home Assistant	`HA_URL`, `HA_TOKEN`
n8n	`N8N_API_URL`, `N8N_API_KEY`
Slack	`SLACK_BOT_TOKEN`, `SLACK_TEAM_ID`
Cloudflare	`CLOUDFLARE_API_TOKEN`, `CLOUDFLARE_ACCOUNT_ID`
Tailscale	`TAILSCALE_API_KEY`, `TAILSCALE_TAILNET`