# FavStash AI Connect — agent router This file is a router. FavStash is a hosted Streamable HTTP MCP server for searching saved Instagram, TikTok, YouTube, and LinkedIn inspiration, analyzing connected social accounts, and planning or publishing content. Do not build or download a local MCP server for ordinary HTTPS-based workflows when the host can connect remotely. Local-file publishing is the separate bridge case below. Hosted endpoint: ```text https://mcp.favstash.app/mcp ``` ## Route to the exact host guide 1. Identify the app or agent running this request. If the user's install prompt names the host, use that name without asking again. 2. Open exactly one matching guide below and follow the selected setup shown at the top of that page. 3. If the host is not listed, use **Other CLI / MCP client** only after checking that host's MCP help or settings. | Detected host | Exact setup guide | | -------------------------------------------- | ---------------------------------------------------------------------------- | | Codex desktop app, CLI, or IDE extension | | | ChatGPT Web, Chat, or Work | | | Claude Web, Desktop, Chat, mobile, or Cowork | | | Claude Code | | | Cursor | | | Windsurf / Cascade | | | Hermes Agent | | | OpenClaw | | | Another CLI or MCP client | | Do not send every guide to the user. Route first, then use the one matching guide. ## Installation authority and permissions The user's one-line install request grants narrow authority to add or update only the `favstash` MCP connection. It does not authorize unrelated config changes. - Prefer the host's built-in MCP command or settings UI over editing a config file by hand. This avoids guessing Windows, macOS, Linux, or WSL paths. - Check the installed host's capabilities before declaring it unsupported. For example, local Codex should check `codex mcp --help`; it does not need a FavStash plugin or connector package to add this remote server. - If the host blocks a required config write, request narrowly scoped approval to run its MCP add command or update its MCP config. Explain the exact file or command before requesting access. - If this session cannot request approval, cannot access the user's local machine, or is running in a cloud-only environment, give the exact command or settings from the routed guide and say that local completion is required. Do not turn a session-level permission limit into the false claim that the host cannot use MCP. - Preserve existing MCP entries. Never overwrite a whole config file to add one server. ## Authentication and secret handling Use OAuth whenever the routed guide supports it. FavStash's OAuth flow opens a browser for sign-in and approval; it does not require an OAuth client ID, client secret, or FavStash API key from the user. An API key is required by the separate local-file bridge, and is an advanced fallback for a client or installed version that cannot complete remote MCP OAuth: 1. Ask the user to create a key at **FavStash → Dashboard → AI Connection**. 2. Ask them to store it in the host's local environment or secret manager under `FAVSTASH_API_KEY`; do not ask them to paste the key into chat. 3. Configure the host to read that variable and send `Authorization: Bearer ` to the hosted endpoint. FavStash keys start with `favstash_pk_`. Never commit, log, print, or echo one. Never request an Instagram, TikTok, YouTube, LinkedIn, Google, or other social-provider password or token; those connections happen inside FavStash. ## Verification contract Do not report success immediately after changing configuration. 1. Use the host's MCP `get`, `list`, `status`, or equivalent to confirm the `favstash` entry exists. 2. Complete OAuth if required, then restart or reload the app, extension, or agent session as directed by the routed guide. 3. Confirm FavStash tools are present. 4. Make a real read-only call, for example: ```text Use FavStash to list my collections and summarize items I saved in the last 7 days. ``` If the host can save the connection but cannot restart itself, tell the user exactly what to restart and describe setup as **configured, awaiting restart**. ## Publishing local media The hosted generic MCP profile currently imports publishing media from an HTTPS URL. It cannot read a local filesystem path. ChatGPT's dedicated profile can translate ChatGPT-native attachments, but that does not make attachment input a generic MCP capability. For a filesystem-capable host, local publishing requires the separately distributed `@sketric/favstash-mcp` stdio bridge. Public npm publication of version 0.3.0 is pending; do not claim an `npx` installation succeeded before the package is available. Use hosted HTTPS media or ChatGPT attachments in the meantime. After installing the bridge, its `create_social_post` and `update_social_post` tools accept absolute `media[].localPath` and `media[].thumbnailPath` values and hide validation, upload, multipart retry, and CDN substitution inside the same tool call. The bridge rejects invalid typed provider settings before reading or uploading local files, and FavStash verifies the stored bytes before returning a content-addressed CDN URL. A remote OAuth-only connection still requires an HTTPS URL unless its host supplies a supported attachment. Do not read or expose the MCP bearer credential, call FavStash's REST upload routes directly, open the web composer, or use unrelated AWS credentials as an upload workaround. ## Reference and troubleshooting - Product and agent overview: - MCP tools: - Public health probe: The `@sketric/favstash-mcp` stdio bridge is for older command-only clients and for filesystem-capable agents that need local-media publishing. HTTP-capable clients using existing HTTPS media should use the hosted endpoint directly.