AgentKavach wraps your LLM client with hard budget enforcement, real-time alerts, and a kill switch that fires before you overspend. One SDK for OpenAI, Anthropic, Google, and Mistral.
Low
Latency overhead
4
LLM providers
4
Alert channels
Three steps to hard budget enforcement. No infrastructure changes required.
One pip install. No infrastructure changes, no sidecar processes.
pip install agentkavachReplace your provider client with AgentKavach. Same API, same parameters, with budget enforcement built in.
guard = AgentKavach(
api_key="ak_prod_...",
llm_key="sk-...",
budget=Budget.daily(50),
on_kill=emergency_stop,
)Every call is tracked. Alerts fire at your thresholds. The kill switch engages at 100%. No surprises on your bill.
response = guard.create(
model="gpt-4o",
messages=[...],
)Put it all together — install, wrap, and ship in a single file:
from agentkavach import AgentKavach, Budget
def emergency_stop():
agent.save_checkpoint()
sys.exit(1)
guard = AgentKavach(
api_key="ak_prod_...",
llm_key="sk-...",
budget=Budget.daily(50),
on_kill=emergency_stop,
)
response = guard.create(
model="gpt-4o",
messages=[{"role": "user", "content": "Hello!"}],
)From a single developer to a fleet of 10,000 agents. AgentKavach scales with your team.
Set daily, monthly, or total budgets per agent. The kill callback fires at 100% to halt runaway spend before it escalates.
OpenAI, Anthropic, Google, Mistral. Track costs across all providers with a single SDK and a unified guard.create() API.
Slack, email, PagerDuty, or custom webhooks. Configure escalating thresholds at 50%, 80%, 95%, and 100%.
See every agent's spend in real time. Drill into model usage, cost trends, and alert history from a single pane.
Token limits, call count caps, runtime limits, and runaway loop detection. Stop agents before they spiral out of control.
Pool a single budget across every agent in your org. Configure via Python SDK using Budget.org_budget(limit, period) or YAML org_budget.
The same library works across OpenAI, Anthropic, Google, and Mistral. Switch providers without changing your budget logic.
from agentkavach import AgentKavach, Budget
# One library, one budget pattern — point a guard at any provider.
def build_agent(provider, llm_key, name):
return AgentKavach(
provider=provider,
llm_key=llm_key,
api_key="ak_prod_...",
agent_name=name,
budget=Budget.daily(50),
)
researcher = build_agent("openai", "sk-...", "researcher")
summarizer = build_agent("google", "AIza...", "summarizer")
# Each provider keeps its native call — AgentKavach tracks spend
# and enforces the budget across all of them.
draft = researcher.create(
model="gpt-4o-mini",
messages=[{"role": "user", "content": "Research today's AI news."}],
)
summary = summarizer.create(
model="gemini-2.5-flash",
contents="Summarize the findings in three bullets.",
)Route every budget threshold to the channel that fits. Start with an email at 50%, escalate to Slack at 80%, page your on-call at 95%, and trigger the kill switch at 100%. Each channel validates its credentials at construction time.
Define budgets, alerts, and shared pools for all your agents in a single configuration file. AgentKavach validates every channel reference, budget type, and shared pool name at load time. Misconfigured alerts fail fast at startup, not when your budget is already up in smoke.
AgentKavach is built with security at every layer. Your API keys, spend data, and agent communications are protected by industry-standard practices from day one. We never use your LLM provider API key for any purpose other than proxying your requests. We only read the response metadata (token counts, model used) to track costs. We never store, log, or access your prompts or completions.
Start free. Scale when you need to.
For solo developers and prototypes.
For teams running production agents.
For fleets of AI agents at scale.
Yes. Upgrade or downgrade at any time. Changes take effect immediately. Downgrades are prorated.
Each LLM API call tracked through the SDK counts as one event. Batch calls count as one event per item in the batch.
Pro comes with a 7-day free trial. No credit card required to start. Max does not include a trial.
Event ingestion pauses until the daily quota resets at midnight in your org timezone (or you upgrade). SDK-side budget enforcement keeps running so agents do not overspend during the pause. The dashboard shows the pause state and resume time.
Max customers can request custom invoicing and payment terms. Contact us for details.