Spend control for your AI agents

One gateway in front of your own OpenAI, Anthropic, Google, Groq and xAI keys. Give every agent its own key, spending cap and kill switch, and stop runaway loops before they burn your budget.

Request access Log in

Works with the OpenAI SDKs, the Anthropic SDKs and Claude Code you already use: change the base URL and the key. Accounts are opened by request during early access; after that, everything is set up from your own dashboard.

# Before: straight to the provider, no limits
client = OpenAI(api_key=OPENAI_KEY)

# After: every agent has its own key and cap
client = OpenAI(
    base_url="https://agentbursar.com/v1",
    api_key=SUPPORT_BOT_KEY,
)
client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[...],
    extra_headers={"X-Agent-Task-Id": "ticket-4182"},
)

What it does

Autonomous agents retry, loop and run unattended. These are the controls that keep that from turning into a surprise bill.

Per-agent keys and hard caps

Each agent gets its own key with daily and monthly caps. At its cap, requests are refused until the window resets, so one agent’s damage is bounded, and you are alerted at 80% and 100% on the way.

Runaway-loop detection

The same prompt repeated in a loop, or a sudden 10× spend spike, throttles that agent to one request a minute and alerts your webhook (Slack-compatible).

A kill switch you control

Suspend any agent from your dashboard. Its next request is refused, whatever it was doing.

Budget downgrades

Past a share of its cap, an agent's requests for an expensive model are served by a cheaper one you choose in your dashboard, instead of stopping outright.

Spend by agent, model and task

Tag requests with a task id and see exactly what each agent, model and job cost, day by day.

Automatic fallback

When a provider has an outage, a request can move to an equivalent model you configured, metered at that model's price. Turn it off per request with one header. Invalid requests come back with the provider's real error.

We never store your prompts or responses. The gateway keeps token counts and costs, not content. Details, including where your requests go, are on the data handling page.

How it works

No SDK to install and no proxy to run.

  1. Point your client at us

    Set the base URL to agentbursar.com/v1 and keep the OpenAI format, streaming included. Claude Code and the Anthropic SDKs use agentbursar.com in Anthropic's own format.

  2. Your provider keys, one key per agent

    In your dashboard, add your provider keys (checked with the provider, encrypted at rest, never shown again; your providers bill you directly) and create a key for each agent with its caps. Name the model you want; it's the model that serves you.

  3. Watch and control

    Your dashboard shows spend per agent and alerts you to runaway behavior. Suspend an agent, change its caps or rotate its key yourself, any time.

Your keys, your provider bill

AgentBursar doesn't resell tokens: you bring your own provider keys and your providers bill you directly. Free during early access. Pricing and provider list prices.

Bring your own keys

Keep your provider accounts. Free during early access; your providers bill you directly.

  • Your traffic uses only your own provider keys
  • Keys encrypted at rest, never shown again after saving
  • Caps, alerts, kill switch and loop detection on every agent

What stays yours

Your provider account, its billing, its rate limits and any dispute with the provider remain between you and your provider.

  • No markup on tokens, no prepaid balance
  • Paid plans later will be a software subscription, with at least 30 days' notice
  • Spend is metered at provider list prices to enforce your caps