MCP
Put Navige in front of an MCP server so its tools are governed.
Navige speaks MCP, so an MCP client — Claude Desktop, Claude Code, Cursor, Windsurf, Zed, an AWS AgentCore Gateway target — can point at it directly. Every tool call that client makes is then identified, recorded and decided before it runs.
Already have local MCP servers in Claude Desktop? navige connect (see Quickstart) puts Navige in front of each of them where they are, so nothing is registered twice: every call is checked before it reaches the server, and a refused call comes back to Claude as a tool error. Claude Code needs no wrapping: its hook already sees every mcp__ tool call.
Configure a client
The key goes in the query string here, because an MCP client configuration is a URL and has nowhere to put a header.
{
"mcpServers": {
"navige": {
"url": "https://api.navige.ai/mcp",
"headers": { "x-navige-key": "nv_your_key_here" }
}
}
}For a client that only speaks stdio, bridge it with mcp-remote:
{
"mcpServers": {
"navige": {
"command": "npx",
"args": [
"-y", "mcp-remote", "https://api.navige.ai/mcp",
"--header", "x-navige-key: nv_your_key_here"
]
}
}
}The gateway still accepts
?api_key= so existing setups keep working, but don’t write new ones that way.What happens on a tool call
- The call is attributed to the agent that key names.
- The kill switch and spend cap apply, on every tier.
- Policies are evaluated. On a paid account a
denystops the call and the client is told why; on Watchtower the decision is recorded and the call proceeds. - An approval rule holds the call until someone answers. MCP clients block on a tool call, so this works without any polling.
Who sees which tools
Every agent in a workspace is offered every tool server connected to it, unless an admin limits a server to some people: Tools & models → the server → Who can use it. A limited server is offered only to agents those people own, and a call to it from anyone else's agent is refused and recorded. An agent belongs to the person who made it, or who connected the computer it runs on. Agents nobody owns, such as ones created through the provisioning API, cannot use a limited server. Use this for a tool that has one person's own account behind it.
Claude Code
Claude Code is a first-class caller. Its workflow keys are derived automatically, so a Claude Code session shows up as real tasks on the spend page rather than as a wall of undifferentiated calls.
AWS AgentCore
An AgentCore Gateway can expose tools over MCP. Point a Gateway Target at https://api.navige.ai/mcp and every tool routed through that gateway is governed, with no code to deploy — the configuration is entirely on the AWS side.