Air-Gapped Deployment

Run Bifrost in environments without outbound internet access.

Overview

Bifrost reaches out to getbifrost.ai for two things:

Data	Default source	Startup behaviour if unreachable
Pricing and model parameter datasheets	https://getbifrost.ai/datasheet	Bifrost cannot start
MCP server library catalog	https://getbifrost.ai/mcp-library	Bifrost starts, MCP Library page stays empty

Every one of these URLs accepts a file:// value, so you can load the data from the local filesystem instead.

---

Datasheets

1. Download the datasheets on a machine with internet access:

curl -o pricing.json https://getbifrost.ai/datasheet
curl -o model-parameters.json https://getbifrost.ai/datasheet/model-parameters

Transfer the files to your air-gapped host (or bake them into your container image / Kubernetes volume).

2. Point Bifrost at the local files in config.json:

{
  "framework": {
    "pricing": {
      "pricing_url": "file:///opt/bifrost/pricing.json",
      "model_parameters_url": "file:///opt/bifrost/model-parameters.json",
      "pricing_sync_interval": 86400
    }
  }
}

An absolute file:// URL takes three slashes followed by the path. Relative references are also accepted (file://./pricing.json or file:./pricing.json) and resolve against the Bifrost process working directory.

3. Ensure the files are accessible to the Bifrost process at the configured paths before starting.

---

MCP server library

The MCP Library page is populated by a catalog synced from https://getbifrost.ai/mcp-library. Air-gapped deployments have two options.

Option A: serve the catalog from a local file

1. Obtain a catalog file. Either download the hosted one on a connected machine, or copy the community catalog that ships in the Bifrost repository at community/mcp-library/servers.json:

curl -o mcp-library.json https://getbifrost.ai/mcp-library

2. Point Bifrost at it in config.json:

{
  "framework": {
    "pricing": {
      "mcp_library_url": "file:///opt/bifrost/mcp-library.json",
      "mcp_library_sync_interval": 86400
    }
  }
}

The same value can be set from the UI under MCP Registry → Library → Settings.

The file is a JSON envelope with a servers array:

{
  "lastUpdatedAt": "2026-06-09T00:00:00Z",
  "servers": [
    {
      "name": "Filesystem",
      "description": "Read and write files on the local filesystem within configured directories.",
      "category": "Developer Tools",
      "connection_type": "stdio",
      "stdio_config": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-filesystem"] },
      "auth_type": "none"
    }
  ]
}

Entries are keyed by a slug derived from name, so re-syncing an updated file updates rows in place. Servers you added yourself through the UI are never overwritten by a sync.

Option B: turn the catalog sync off

If you do not want a server catalog at all, set the interval to 0:

{
  "framework": {
    "pricing": {
      "mcp_library_sync_interval": 0
    }
  }
}

Bifrost then skips the catalog fetch at startup and never schedules a background sync, so no requests go to getbifrost.ai. Anything already stored in the database is still served, MCP servers can still be added manually, and Force Sync Now in the UI still runs on demand.

Catalog entries carry an icon_url pointing at a remote image. On an air-gapped host those images do not load and the UI falls back to a generic MCP icon. This is cosmetic and does not affect server connectivity.

---

Keeping local data current

Bifrost re-reads each local file on every sync tick, so updating a file on disk is enough. No restart is needed.

Setting	Cadence	Minimum	Disable
pricing_sync_interval	Default 24 h	3600 s	Not supported
mcp_library_sync_interval	Default 24 h	3600 s	Set to 0