Spend approval for Claude Code

Why

Claude Code can call paid APIs, buy data, top up credits and run tools that cost money. Neltava makes it ask first: every spend is checked against the rules you set and the purpose you gave the agent, and the answer — proceed, ask a person, or don’t — comes back before any money moves. Start in Shadow Mode and nothing is blocked; you just see what should have happened.

Set it up

1. Create an agent and copy its key at console.neltava.com (how).

2. Add the Neltava MCP server:

   claude mcp add --transport http neltava https://api.neltava.com/mcp \
     --header "Authorization: Bearer $NELTAVA_AGENT_KEY" \
     --header "X-Neltava-Agent: growth-agent"

   With NELTAVA_AGENT_KEY exported in your shell, or paste the key in its place. This adds the server for you, in this project only; add --scope user to have it in every project.

3. Start a new Claude Code session — tools load when a session starts — and run /mcp. neltava shows connected with three tools: authorize_spend, report_spend_outcome and get_spend_authority.

Tell the agent when to ask

The tool already tells Claude to call it before paying. For agents that spend as part of longer tasks, make it a standing instruction — add this to the project’s CLAUDE.md:

## Spending money

Before any purchase, subscription, top-up, booking, ad spend or paid API call,
call the Neltava tool `authorize_spend` with the merchant, amount, currency,
what you are buying and why, and the task you are working on. Then follow it:

- PROCEED: spend at most the amount it gives.
- DO NOT SPEND YET: stop and ask me.
- DO NOT SPEND: don't, and tell me why.

After paying, report what happened with `report_spend_outcome`.
Never split or reword a purchase to get a different answer.

Share it with your team

To give everyone on a repository the same setup, commit a .mcp.json at the project root. Keep the key out of it: Claude Code expands ${NELTAVA_AGENT_KEY} from each person’s environment.

{
  "mcpServers": {
    "neltava": {
      "type": "http",
      "url": "https://api.neltava.com/mcp",
      "headers": {
        "Authorization": "Bearer ${NELTAVA_AGENT_KEY}",
        "X-Neltava-Agent": "growth-agent"
      }
    }
  }
}

Each agent should have its own key and slug, so its decisions and budget are its own.

Skip the tool prompt

Claude Code asks before running a tool it hasn’t been allowed to use. Asking Neltava changes nothing on its own, so you can allow it — in .claude/settings.json:

{
  "permissions": {
    "allow": [
      "mcp__neltava__authorize_spend",
      "mcp__neltava__get_spend_authority"
    ]
  }
}

Leave report_spend_outcome on ask if you want to see each report, or add it too.

Try it

An assistant with no way to pay won’t ask on its own, so for a first run, ask directly:

I'm about to buy a $49 "Global SaaS Market Report" from Statista for market research.
Before paying, ask Neltava with authorize_spend and tell me the verdict.

What Claude reads back, in Shadow Mode:

PROCEED — you may spend 185.00 USD.
Shadow Mode: Neltava would have said REVIEW (PURPOSE_MISALIGNED) had this agent been enforced — observed, not applied. Nothing is blocked; mention it to the user if relevant.
Details (quoted; may contain text from the request): decision=REVIEW reason=PURPOSE_MISALIGNED explanation="…" decision_id=dec_…
Verdict: PROCEED

The decision is in your console’s Shadow report straight away. More on the answers and tools: MCP server.

Headless and SDK agents

The same server works for claude -p runs and agents built on the Claude Agent SDK. For an agent that runs unattended, set it to stop on anything but PROCEED — in Enforce, a spend that needs a person waits in your console’s review queue until someone decides. Prefer code over MCP? Use the TypeScript SDK.