Catalogs

Reference for the remote MCP server and on-device model catalogs, their schemas, and how to contribute.

Mobile Agent maintains two remote catalogs in the repository, fetched at runtime and cached in memory for 30 minutes (with a bundled copy as offline fallback):

catalog/mcp-servers.json — preset MCP servers.
catalog/on-device-models.json — on-device LiteRT-LM models.

When a catalog change is merged to main, users receive it when the cache expires or the app restarts — no app rebuild is required.

Schema note: the top-level version field is the catalog schema version, not a content revision. Keep it at 1 when adding entries. Incrementing it requires shipping parser support in the app first; older releases reject unknown versions.

`mcp-servers.json`

jsonc

{
  "version": 1,
  "servers": [
    {
      "id": "notion",
      "label": "Notion",
      "description": "Search, read, and update pages in your Notion workspace.",
      "url": "https://mcp.notion.com/mcp",
      "transport": "http",
      "authMode": "oauth",
      "oauthMode": "proxy"
    }
  ]
}

Field reference:

Field	Type	Required	Notes
`id`	string	yes	Stable, unique catalog identifier.
`label`	string	yes	Name shown in the app.
`description`	string	yes	Short explanation shown below the name.
`url`	string	yes	Public HTTPS MCP endpoint.
`transport`	`"http" \| "sse"`	yes	MCP transport.
`authMode`	`"none" \| "oauth" \| "headers"`	yes	Authentication mode.
`headerTemplate`	string	no	Public placeholder shown in the headers field (e.g. `Authorization: Bearer <…>`).
`oauthClientId`, `oauthAuthorizationUrl`, `oauthTokenUrl`, `oauthScopes`, `oauthAllowedAuthOrigin`	string	no	Optional public OAuth overrides.

Security: never add API keys, access tokens, client secrets, or private headers to this file. User credentials remain in the app's encrypted on-device secret store.

`on-device-models.json`

jsonc

{
  "version": 1,
  "models": [
    {
      "id": "qwen3-4b",
      "name": "Qwen3 4B",
      "parameterCount": "4B",
      "quantization": "mixed-int4",
      "downloadUrl": "https://huggingface.co/litert-community/Qwen3-4B/resolve/main/qwen3_4b_mixed_int4.litertlm",
      "sha256": "f0794bc77efeaaf4f7af815f04c483b19b8f2ae4a102cef1b7b760a25848a18e",
      "sizeBytes": 2659057664,
      "contextWindow": 2048,
      "lowMemoryContextWindow": 2048,
      "minRamBytes": 8000000000,
      "minRamBasis": "benchmark-derived",
      "supportedPlatforms": ["ios", "android"],
      "license": "Apache-2.0",
      "capabilities": { "tools": false, "reasoning": false }
    }
  ]
}

Only LiteRT-LM compatible files (.litertlm) should be added. Field reference:

Field	Type	Required	Notes
`id`	string	yes	Stable, unique lowercase identifier. Do not reuse an old ID.
`name`	string	yes	Name shown in Settings.
`parameterCount`	string	yes	Short display label (e.g. `4B`).
`quantization`	string	yes	Quantization display label.
`downloadUrl`	string	yes	Public HTTPS URL to the model file.
`sha256`	string	yes	Exact 64-char hash for integrity verification.
`sizeBytes`	number	yes	Exact file size, used by progress + validation.
`contextWindow`	number	yes	Practical token limit for the packaged model.
`lowMemoryContextWindow`	number	no	Context window used on low-memory devices.
`minRamBytes`	number	yes	Minimum device RAM to enable the model.
`minRamBasis`	string	no	`published-minimum`, `benchmark-derived`, or `analogous-model-derived`.
`backend`	`"cpu" \| "gpu"`	no	Inference backend.
`supportedPlatforms`	`("ios" \| "android")[]`	yes	Platforms that can run it.
`license`	string	yes	Model-weight license shown to the user.
`capabilities.tools`	boolean	yes	Whether the model can call tools.
`capabilities.reasoning`	boolean	yes	Whether the model supports reasoning.

Contribution checklist

Before merging a catalog change:

Test the exact URL, size, hash, and device memory requirement.
Changing a URL without updating the pinned size/hash makes the download fail safely (which is correct, but test it).
Never include secrets.
Keep version: 1.