Your AI agents spend money. Give each one a spending limit.

Try it with your own numbers. No signup, nothing stored — this calls the real check the proxy runs, with the numbers you enter.

Cost metadata only. An entry stores the agent id, rail, service label, amount, token counts, model and timestamp. There is no field for prompt or response content and no code path that writes it — a test asserts it stays that way. Read the security page →

Your provider key is yours. The proxy is pass-through: the key rides in the request headers, is forwarded verbatim, and is never stored — a test greps the whole data directory for it after a proxied call and fails if it appears. You can use the proxy without giving us a key at all.

Claude Code

claude mcp add --transport http agent-ledger https://aiagentscity.com/mcp/

Codex

codex mcp add agent-ledger --url https://aiagentscity.com/mcp/

Cursor / any MCP client — merge into mcp.json:

{
  "mcpServers": {
    "agent-ledger": {
      "url": "https://aiagentscity.com/mcp/"
    }
  }
}

Or the Python wrapper — one line, every call metered and capped:

pip install "aiagentscity-ledger[wrapper]"
client = agentledger.wrap(OpenAI(), agent_id="YOUR_AGENT_ID", agent_secret="as_...")

Published on PyPI as aiagentscity-ledger — the umbrella namespace, which nobody else can register because it is our domain. agent_secret comes from agent-ledger init --agent YOUR_AGENT_ID (printed once, also written to .env), or from the response to a first POST /v1/track that carries your workspace_key in the body.

Tools: ledger_track, ledger_set_budget, ledger_check_spend, ledger_report, ledger_alerts, ledger_price, ledger_proxy_attach, spend-approval tools ledger_request_approval, ledger_approvals and ledger_approval_decide, plus open docs tools ledger_api_docs and ledger_examples. Registered as io.github.entradox/agent-ledger.

Try it in your agent: "Track my Claude Code spend" · "Alert when any agent exceeds $50/day" · "Weekly P&L report".

Two mechanisms, two different guarantees — and we never blur them.

Through the proxy or the wrapper: the call is refused with 402 before the provider is contacted, so the money is never spent.

On the plain API: the cap can only block the write when crossed — the ledger entry that would cross it is rejected, but the provider has already charged. That stops the recording, not the spend.

Traffic that bypasses the proxy is not enforced. Any caller that can reach your provider directly is outside our reach. If a cap has to be real, route the traffic through the proxy or the wrapper.

Trace tools answer "what did my agent do?" — $39/seat (LangSmith), $79/mo (Helicone), $249/mo (Braintrust). They bill per seat or per trace, and none of them stop the $401st call. AgentLedger is free for up to 3 agents per workspace, and the agent itself can buy more over x402 — no human in the loop. Full comparison, with dated prices →

1. Get a workspace_key — no signup, no login, no card: /start. Running as an agent with a wallet, with no human at all: POST /v1/billing/x402 — the paying wallet becomes the workspace identity. settles on Base MAINNET (eip155:8453) in real USDC. The paying wallet pays real money; the workspace is live immediately.

2. Track a spend event — dollars and tokens

curl -sL --post301 -X POST https://aiagentscity.com/v1/track \
  -H "Content-Type: application/json" \
  -H "AL-API-Version: 2026-09-01" \
  -d '{"agent_id":"YOUR_AGENT_ID","rail":"x402","amount_cents":100,"service":"search","tokens_in":4500,"tokens_out":1200,"model":"gpt-4o","workspace_key":"YOUR_WORKSPACE_KEY"}'

A key claims, a secret writes: the first call for a new agent_id must carry "workspace_key":"wk_live_..." in the body — that claims it into your workspace — and returns that agent's agent_secret. Save the secret: every later write to the same agent_id sends "agent_secret":"..." instead and needs no workspace_key. Reads (report/tokens/alerts) require an X-Agent-Secret or X-Workspace-Key header — over REST and MCP alike.

The AL-API-Version header is required on writes. Omit it and the call is rejected — that is deliberate, so a future breaking change cannot silently alter what your agent meant.

3. Set a monthly cap — enforced, not just logged:

curl -X POST https://aiagentscity.com/v1/budget \
  -H "Content-Type: application/json" \
  -H "AL-API-Version: 2026-09-01" \
  -d '{"agent_id":"YOUR_AGENT_ID","monthly_cents":5000,"agent_secret":"YOUR_SAVED_SECRET"}'

4. Pull the P&L — totals, by rail, by service, budget status and anomalies:

curl "https://aiagentscity.com/v1/report/YOUR_AGENT_ID?days=30" \
  -H "X-Agent-Secret: YOUR_SAVED_SECRET"

Token burn: /v1/tokens/YOUR_AGENT_ID. CSV: /v1/report/YOUR_AGENT_ID/csv. Browser view: /v1/report/YOUR_AGENT_ID/html. Everything at once: /dashboard.