Skip to main content

MCP reference

The tools, what they cost, host behaviour on long tasks, and the things that catch people out.

The visual install guide for Claude, Cursor, ChatGPT and Manus is at getvamos.ai/mcp. This article is the reference: tools, limits, host behaviour, billing and troubleshooting.

​

​

The tools

Your AI client picks the right one from what you ask. You do not call them directly.

  • chat_completion Runs a full Vamos turn and waits for it. Only for hosts that cannot poll, because hosts that proxy MCP drop a call that runs for minutes.

    Spends credits: Yes

  • chat_completion_start Starts a turn in the background and returns a run id at once. The default on every host, because hosts that proxy MCP drop long synchronous calls.

    Spends credits: Yes, on completion

  • chat_completion_poll Checks a run. Long-polls, so it returns as soon as the status changes.

    Spends credits: No

  • chat_completion_cancel Aborts a run so credits stop accruing.

    Spends credits: No

  • create_chat Starts a new Vamos conversation and returns its id.

    Spends credits: No

  • list_chats Lists your chats, so the AI can continue one rather than starting fresh.

    Spends credits: No

  • get_messages Reads a chat's history. Also how to check whether a timed-out turn already finished.

    Spends credits: No

  • get_credit_usage Your remaining credits and plan.

    Spends credits: No

  • auth_status Whether the session is authenticated, and the sign-in URL if not.

    Spends credits: No

​

A completion has the same reach as the web app: video search, web search, profile and post analysis, comment pulls. It writes using your voice and your memories from My Brand.

​

What search covers: Instagram, TikTok and YouTube. Ask and it will also go after YouTube long-form and Facebook Ads. LinkedIn, X and Facebook can be read from a link but cannot be searched.

​

​

The using-vamos skill

For Claude Code, install the using-vamos skill alongside the MCP. It teaches Claude how to call Vamos well rather than merely that it can.

​

Download it from getvamos.ai/skill, put it at ~/.claude/skills/using-vamos/SKILL.md, and restart Claude Code. It activates when you ask about social media.

​

What it adds over a bare MCP connection:

  • Continues an existing chat rather than starting a new one each time.

  • Picks the background path for anything substantial.

  • Folds a multi-part request into one completion instead of fanning out.

  • Passes Vamos's answer through rather than rewriting it.

  • On a timeout, checks whether the turn already finished before retrying.

​

The skill is Claude Code only for now.

​

​

What you cannot do

No file upload and no doc editing. Those stay in the web app.

​

No streaming. You get the full answer when Vamos finishes, which is anywhere from a few seconds to 15 minutes.

​

No raw tool results. You get Vamos's written answer. The underlying search results, transcripts and comment dumps are in the chat in the web app.

​

No switching workspace mid-session. The token is bound to the workspace you were in at sign-in. Switch in the app, then sign out and back in inside your client.

​

Retries are not free. Re-running a completion runs the whole job again and bills again. On a timeout, call get_messages first.

​

​

Host behaviour on long tasks

Client

Live progress

What you see

Claude Code

Yes

Progress mid-call. Best experience, especially with the skill.

Claude Desktop

Yes

Progress mid-call.

Cursor

Minimal

Little until it is done.

ChatGPT

No

Silent until the result returns.

Manus and other agents

Varies

May retry hard, and each retry bills. Cap retries at one.

Codex

Partial

Depends on the version.

On ChatGPT and Cursor, expect up to 15 minutes of silence on heavy research. Break big requests into smaller ones.

​

​

Billing

MCP runs on the same credits as the web app and costs no extra.

​

Only chat_completion and chat_completion_start spend. Everything else is free. A background run bills when it completes even if you stop polling, so cancel rather than abandoning it.

​

Out of credits returns ERR_INSUFFICIENT_CREDITS. On an active paid plan, buy extra credits in Settings, under Billing. A locked workspace or an ended plan returns ERR_WORKSPACE_LOCKED, even with credits left.

​

​

Troubleshooting

The AI is writing for the wrong workspace. The token is bound to the workspace from sign-in. Sign out of the Vamos MCP in your client and reconnect.

​

Nothing seems to be happening in Cursor. Cursor does not show tool progress. Give it 15 minutes, then check your credit balance.

​

My agent is burning credits. Some agents retry failed calls, and each retry re-runs the completion. Set a retry cap of one.

​

The skill will not activate. Check the path is exactly ~/.claude/skills/using-vamos/SKILL.md, restart, and ask something explicitly about social media.

​

The AI told me to open a dashboard page that does not exist. It hit an auth error and invented navigation. Ask it to call auth_status, which returns the real sign-in URL, or reconnect in your client's connector settings.

​

Can I revoke access? Sign out of the MCP server in your client, or click Disconnect beside it in Vamos, under Settings, then Connectors.

​

Does it work for a team? Each person authenticates with their own account. There is no shared key.

​

Which plan do I need? Any plan with credits, Free included. A locked workspace or an ended plan is refused.

​

​

Related

Did this answer your question?