convez.ai
MCP Server
MukundaKatta

AgentBudgetMcp

MukundaKatta/AgentBudgetMcp

Ver no GitHub
InicialPluginScore · de 0 a 100

MCP server for @mukundakatta/agentbudget — token + dollar budget caps for AI agents over stdio. Drop into Claude Desktop, Cursor, Cline, Windsurf, Zed.

Instalação

$npx convez-ai MukundaKatta/AgentBudgetMcp

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 2 meses
Versão
v0.1.0
Issues abertas
4
Contribuidores
1

README

A tradução para o português deste README ainda está na fila.

agentbudget-mcp

MCP server for @mukundakatta/agentbudget. Exposes token + dollar budget caps to any MCP-aware client (Claude Desktop, Cursor, Cline, Windsurf, Zed) so agents can track spend across tool calls and refuse work that would push past the ceiling.

Install

The cleanest path is to let your MCP client launch the server via npx:

Claude Desktop — claude_desktop_config.json:

{
  "mcpServers": {
    "agentbudget": {
      "command": "npx",
      "args": ["-y", "@mukundakatta/agentbudget-mcp"]
    }
  }
}

Cursor — ~/.cursor/mcp.json or <project>/.cursor/mcp.json:

{
  "mcpServers": {
    "agentbudget": {
      "command": "npx",
      "args": ["-y", "@mukundakatta/agentbudget-mcp"]
    }
  }
}

Same shape works for Cline, Windsurf, Zed — point at the @mukundakatta/agentbudget-mcp package over stdio.

Tools

The server exposes three tools that share one in-memory budget for the duration of the MCP session:

configure_budget

Set the active caps. Calling this replaces any existing budget; totals reset to zero.

ArgumentTypeNotes
max_input_tokensnumberoptional — cap on cumulative input tokens
max_output_tokensnumberoptional — cap on cumulative output tokens
max_total_tokensnumberoptional — cap on input + output combined
max_cost_usdnumberoptional — cap on dollars; needs pricing
pricingobjectoptional — per-model { inputPer1k, outputPer1k } map; falls back to the built-in starter table
allow_unknown_pricingbooleanoptional — when true, models with no pricing cost $0 instead of erroring

record_llm_usage

Tally one LLM call. Call this after the response — by the time you have token counts, the call already cost money.

ArgumentTypeNotes
modelstringrequired when max_cost_usd is set
input_tokensnumberrequired
output_tokensnumberrequired

Returns one of:

  • { "status": "ok", "totals": {...}, "remaining": {...} } — call accepted
  • { "status": "exceeded", "cap": "...", "limit": ..., "attempted": ..., "overshoot": ..., "totals": {...} } — a cap tripped. This is data, not an MCP error — your agent can read it and decide what to do (fall back, summarize, stop).

budget_status

Peek at current totals + per-cap remaining. Useful for letting the agent decide whether to schedule more work.

Why an MCP server?

You can use @mukundakatta/agentbudget (or its Python sibling) directly inside your own agent code — that's the lower-friction path.

This MCP wrapper exists for two cases:

  1. Cross-tool budgets: when several MCP-spawned tools share one session, this server keeps a shared spend counter that no individual tool owns. Useful for caps like "do not exceed $5 across this whole session, regardless of which sub-agent runs."
  2. Client-driven enforcement: when the client (Claude Desktop, Cursor) is the one driving model calls and you want a budget surface the operator can configure declaratively without modifying agent code.

Otherwise, prefer importing the lib directly.

Sibling libraries

Part of the @mukundakatta/agent* reliability stack:

Direct lib: @mukundakatta/agentbudget (npm) · agentbudget-py (PyPI)

License

MIT © Mukunda Katta

Ver o original no GitHub

Bom para

Relacionados