TokPortal MCP Server: Setup, GitHub, npm and Tools

Technical review: September 18, 2026, against the public documentation, current API schema and published npm package 1.15.1. Examples below are instructional templates, not customer results or a claim that a paid campaign was executed.

If you arrived from the TikTok account-creation MCP announcement, start here. This page gives you the actual endpoint, official repository, installation choices and a way to find the tool you need. You do not need to build a publishing integration before trying one useful operation.

The first win is simple: your assistant reads your available countries and balance from TokPortal. The next is a costed preview for one account. Once those work, move to a complete publishing workflow. A working connection and a delivered TikTok post are different milestones; this guide helps you tell them apart.

MCP stands for Model Context Protocol. It lets an AI host discover structured tools, read their input schemas and call a connected service. TokPortal supplies the account and publishing operations; your assistant turns your request into tool calls.

In the founder’s follow-up answers, the operating model is described as a marketplace of real local community managers. Account creation and native posting are human work coordinated through software. Connecting an AI assistant does not turn the service into an emulated-phone script or make delivery instantaneous.

This is useful when you have approved content and need a repeatable way to request accounts, configure publishing, track delivery and inspect results. It is also useful for agencies building a service around that work. It does not decide whether your content is interesting or promise that a localized account will reach only local viewers.

The remote Streamable HTTP endpoint is https://app.tokportal.com/api/ext/mcp. Use tokportal/tokportal-mcp on GitHub for source and issues, and tokportal-mcp on npm for the local server. The developer guide documents both transports.

The package checked for this guide is 1.15.1, with 91 management tools and a Node.js 20-or-newer requirement. Your connected server’s tool discovery is the authority for its current catalog. Do not infer that every newly documented API capability is already included in an older local package.

Prefer the remote connection when you want browser sign-in and no local runtime to maintain. Use stdio when your client requires a local process or your workflow needs local-file uploads. Both routes authenticate to your TokPortal workspace.

Choose your client below and merge the snippet into its existing configuration. Do not replace unrelated MCP servers. The remote examples use OAuth, so you complete authentication in the browser instead of copying a secret into your conversation.

For Claude’s custom-connector route, follow the TokPortal remote connection guide: add the endpoint, sign in and choose an access level. In ChatGPT, use custom remote MCP/app setup where it is available in your workspace, select OAuth and enable the connection for the conversation. Menu names and availability depend on the host and account.

For Codex specifically, OpenAI’s MCP documentation confirms remote HTTP servers, OAuth login and the mcp_servers configuration. After adding the TOML section, run codex mcp login tokportal, then inspect codex mcp list. A configuration file existing on disk is not proof that the running client loaded it.

Choose read-only access for balance, account and analytics checks. Account creation and publishing require write access. A dry-run write still goes through authorization, so read-only access does not become write access just because the operation is simulated.

Ask your assistant to use tokportal_get_credit_balance. Inspect the actual tool result: you want a successful response from your workspace, not a paragraph describing what your balance might be. Follow with tokportal_list_countries and tokportal_get_credit_costs when planning an order.

If no tool is called, confirm that the server is enabled in this conversation and the host loaded its tools. If authentication fails, finish OAuth or check the private key configuration. The troubleshooting guide separates connection, permission and request errors.

For a new account, the management flow starts with tokportal_create_bundle, then tokportal_configure_bundle_account. One bundle represents one account. Upload and configure approved media, inspect tokportal_get_bundle_publish_readiness, then submit the ready work. Read the returned state and follow delivery separately.

For an already delivered managed account, use a videos_only bundle tied to that account’s ID. For reporting, start with the analytics dashboard or account-video analytics tools available to your workspace. For warming, inspect the current advanced-warming parameters and the returned session evidence. The explorer above lets you search these exact names without executing them.

Tool names are a map, not a substitute for input schemas. JSON request fields go under body; path arguments such as id and position stay at the top level. The tutorials below show that envelope explicitly so your assistant does not invent a flat request.

For supported management writes, send top-level dry_run: true using the remote endpoint or local package 1.15.1 or newer. Check for dry_run: true, credits_charged: 0 and the returned credits_would_charge. Review that quote before a live operation.

Do not use local 1.15.0 for simulations. The official docs identify a version-specific issue: it advertised the argument but did not forward the dry-run header. Update the configured package and restart the server. If you cannot establish the running version, use the remote endpoint.

Live bundle creation debits credits immediately, even while its status is pending_setup. Unpublishing is not a refund. Include ongoing managed-account coverage in your budget, and use the current quote rather than a copied price from a launch post. Synthetic IDs from a simulation cannot be used in live requests.

The launch FAQ mentions an additional generation mode. The current public API schema describes AI model discovery, quotes, generation jobs and a separate USD wallet with permission checks. That is a different contract from the management-credit examples in this guide.

Discover the models and tools actually available to your workspace, obtain the required quote and use the documented maximum-cost approval flow. Do not apply management dry_run assumptions to AI generation: those endpoints do not advertise that legacy contract. An installed management package alone does not prove you have access to a particular generation model.

Account-Generated Distribution (AGD) means organizing distribution around purpose-built accounts with clear audiences and recurring formats. MCP supplies an operating interface for the account network. The creative brief still explains who each account serves and why someone would follow it.

For example, a language app could run a travel-Spanish series and a workplace-Spanish series, using its own product demonstrations and distinct scripts. Track which content gets delivered and what visitors do afterward. That is a workable test; copying the same asset to a hundred profiles tells you much less about the audience.

The founder confirms white-label compatibility in the FAQ. If you plan to resell, define the customer offer, approval process and support boundaries first. Read the white-label API guide and reseller page before packaging recurring operations as “unlimited.”

MCP practical guides: installation, GitHub and tools · create TikTok accounts · Instagram accounts and publishing · TikTok distribution · fix MCP errors.