Docs / MCP

Let your agent run the tab.

One URL. Any MCP client reads spend, closes a tab, changes a cap, as you and only as far as your role allows.

Server

https://console.flocktab.com/api/mcp

Streamable HTTP with OAuth. Your client opens the console; you approve; nothing to paste. Agents can read this page as plain markdown at /docs/mcp.md.

Set up

  1. Claude Code

    Add the server, then run /mcp inside Claude Code, pick flocktab and Authenticate.

    claude mcp add --transport http flocktab https://console.flocktab.com/api/mcp
  2. Codex

    Add the server, then sign in once; the browser opens the console.

    codex mcp add flocktab --url https://console.flocktab.com/api/mcp
    codex mcp login flocktab
  3. Cursor

    In .cursor/mcp.json (or ~/.cursor/mcp.json for every project). Cursor asks you to connect the first time.

    {
      "mcpServers": {
        "flocktab": { "url": "https://console.flocktab.com/api/mcp" }
      }
    }
  4. claude.ai and Claude Desktop

    Settings, Connectors, Add custom connector. Paste the URL and connect.

    https://console.flocktab.com/api/mcp
  5. Any other MCP client

    Streamable HTTP with OAuth. Point it at the URL; it discovers the rest.

    https://console.flocktab.com/api/mcp

Access

  • The agent acts as the person who approved it, in the flock they chose, with their role's permissions. Project-limited roles stay limited.
  • Permissions are checked again on every call, so a role change or suspension applies at once.
  • Nothing to paste: OAuth 2.1 with PKCE and dynamic client registration, like Resend's and Cloudflare's MCP servers. The console's sign-in, passkey and SSO rules apply to the approval.
  • Access tokens last an hour and refresh themselves for 30 days. Account, Security, Connected apps lists every connection; Disconnect revokes it at once.

Tools

ToolWhat it doesNeeds
whoamiThe person and flock this connection acts for, and what their role allows.signed in
list_tabsEvery Agent's tab: open or closed, cap, spend this window, project.spend:read
get_spendSpend for month, 7d, 30d or 90d: metered, outside, subscription use, by project, Agent and day.spend:read
list_decisionsLedger rows, newest first: each call the proxy allowed or refused, with its amount and reason, and total, every row in the ledger. For older rows pass next_before back as before; it pages to the first row without repeating one.spend:read
list_guard_holdsWhat Agent Guard held for a person and what people decided, as the Agent Guard page shows it: Waiting, Blocked, Allowed or All, each tab's count (all time), newest first with older pages by cursor; and this month's Agent Guard judgments against the plan's cap with its reset in the flock's time. Read only: it cannot allow or refuse a call. Team and Fleet.spend:read
close_tabStop one Agent. Its next call is denied before the provider.tabs:write
reopen_tabLet a closed Agent spend again.tabs:write
set_tab_capChange an Agent's cap, in dollars.tabs:write
close_all_tabsThe kill switch: close every tab in the flock, including any its cap would reopen at the next window. Returns how many it closed.tabs:write, whole flock
open_all_tabsReopen every closed tab in the flock, except one set to close for good at its cap. Returns how many it reopened.tabs:write, whole flock
list_outside_spendBills that never went through the tab (synced vendors and bills recorded by hand), for a range, one day or one vendor: that set's total and per-project totals, the largest lines (up to 500), and how many smaller lines were left out. A line split across projects reads project "shared" and lists each project's part; a $0 line inside a vendor allowance shows its usage.outside:read
record_outside_spendAdd spend that bypassed the tab (an invoice, a contract, a card charge). Team and Fleet.outside:write
attribute_outside_spendGive outside spend to a project, share it by Agent activity, or clear it.outside:write, whole flock
  • Money is dollars as strings with cents, for example "12.34".
  • Tools name Agents by slug, as list_tabs returns them.
  • A tool the role does not allow answers with an error saying so. Do not retry it; ask the person.