AgentWach is the AI agent budget firewall.
Track every AI agent, GitHub Copilot credit, token, tool call, and dollar in real time — then warn, pause, or stop agents before they burn your budget.
No credit card required · 2-minute setup · connects to anything that calls an LLM
Live numbers from the agentwach network · refreshed every 60 seconds
Built for the people running AI agents.
From a solo builder with one agent to an enterprise running thousands — pick the playbook closest to yours.
Solo AI builders
Ship side projects without waking up to a $400 OpenAI bill. Budgets, alerts, and an emergency stop in 2 minutes.
AI agencies
Watch every client agent in one room. Per-workspace caps, white-label reports, audit-ready logs.
GitHub Copilot teams
Track Copilot premium-request credits by user, repo, and model. Cap a repo before it hits 100%.
LangGraph / CrewAI / AutoGen devs
Drop-in tracing for multi-agent graphs. Catch loops between nodes before they burn the whole run.
Multi-provider teams
OpenAI, Anthropic, Gemini, and Copilot in one dashboard. One budget, one alert channel, one source of truth.
Enterprise AI governance
Org-wide caps, SSO, audit logs, and prompt-injection blocking across every provider your teams use.
One bad agent loop can run all night.
Most tools tell you after. AgentWach tells you before — and can pull the plug.
Runaway Loop Detected
Agent repeats the same tool call hundreds of times. AgentWach hard-caps it and returns stop=true.
Budget Spike Alert
Research bot is burning through its daily budget faster than usual. Live alarm fires at your warn threshold.
Risky Prompt Blocked
Possible prompt injection scanned against 20+ rules. Optional Gemini verification confirms and suggests a fix.
AgentWach does not just show spend. It shows avoided spend.
Every guardrail stop is a dollar that stayed in your budget. Here is what real protection looks like.
412 repeat calls caught in 6 minutes before the loop compounded.
Daily cap alarm fired at 82% of budget. Agent paused before overspend.
Usage jumped 6× on one repo. Alert fired before credits disappeared.
Malicious ignore-previous rule scored 0.92. Request killed instantly.
Same tool invoked in a tight loop. Hard stop returned before the bill grew.
Adjust the sliders to match your agent's behavior. See what an unchecked loop would burn in a single workday.
Don't just observe agents. Control them.
Competitors ship graphs and traces. AgentWach ships buttons. Warn, pause, throttle, require approval, block, hard-stop — every guardrail is one toggle away.
From silent loop to stopped agent in under a minute.
A real runaway-agent scenario, end to end.
Multiple agents running
research_bot, support_triage, lead_enricher, rogue_scraper — all live.
One agent starts looping
rogue_scraper hits the same scrape() tool 40 times in 20 seconds.
Repeat tool-call pattern detected
AgentWach groups the calls and flags the loop signature.
Budget alert fires
$30 daily cap hit 94%. Slack + email pinged in under a second.
AgentWach stops the agent
Hard cap returns stop=true. rogue_scraper halts mid-run, no manual intervention.
Dashboard shows money saved
$218 protected. Replay tab shows the exact tool calls that triggered the stop.
From signup to safety net in three steps.
Connect your agents
SDK, REST, OTLP, CSV, email, or browser extension. First event lands in under a minute.
Set budgets & guardrails
Per-agent token and dollar caps, warn thresholds, re-alert cadence, loop protection — all in the UI.
Catch loops & runaway spend
Live alarms fire instantly. Hard caps return stop=true so agents self-halt before damage spreads.
Every agent, every token, every dollar.
A real-time monitoring room for your AI stack.
Play with AgentWach before you connect anything.
Runaway Agent Simulator
Trigger a loop and watch AgentWach stop it in real time.
Run simulationPrompt Injection Scanner
Paste any prompt. We scan it against 20+ injection rules.
Scan a promptToken Burn Calculator
Model your monthly token spend across providers.
Open calculatorGitHub Copilot Credit Estimator
Estimate Copilot premium-request credits before your team blows the quota.
Estimate creditsSample Client Report
See exactly what an end-of-month agent report looks like.
View reportHow agentwach saves you money.
A single runaway loop can quietly drain a budget overnight. Drag the sliders to model your stack — see what one bad agent would cost without a circuit breaker, and what it costs when agentwach trips the alarm.
Estimate only. Actual savings depend on your provider pricing, prompt size, and how fast you wire alarms to a kill switch. agentwach Free includes the alarm engine.
If it calls an LLM, AgentWach can watch it.
ChatGPT
Browser extension · OpenAI key polling
Claude
Browser extension · Anthropic key polling
Gemini
Browser extension · API polling
GitHub Copilot
Premium-request credit tracking
Codex / OpenAI Codex
Per-task token + cost capture
LangGraph
Per-node tracing · loop detection between nodes
CrewAI
Per-crew, per-agent token + cost rollup
AutoGen
Multi-agent group-chat traces
Vercel AI SDK
Drop-in middleware for AI SDK route handlers
OpenTelemetry
OTLP/HTTP endpoint · any OTel exporter
REST API
POST events to /api/public/ingest
CSV upload
Drop usage exports · auto-mapped columns
Email forwarding
Forward provider receipts to a private inbox
Browser extension
Pair with a 6-char code · Chrome & Firefox
Send guardrail stops where your team already works.
Route by severity, agent, or workspace. Suppress noise, escalate the real ones.
Slack
Per-channel routing · @mentions on hard stop
Discord
Webhook into any server
Microsoft Teams
Incoming webhook · per-channel
PagerDuty
Page on-call for critical guardrail stops
Threaded per-incident · re-alert cadence
Webhook
JSON POST to any endpoint
SMS
Critical-only · so 3 a.m. means 3 a.m.
Built for teams shipping with
We collect the bare minimum to keep your agents safe.
No prompt capture by default. No customer PII. Honest about what's done and what's in progress.
SOC 2 Type II
Audit in progress · controls aligned.
GDPR-ready
DPA available · EU-region hosting available.
API keys hashed
Stored as SHA-256 · raw keys never persisted.
Prompt capture optional
Off by default · you choose what we see.
EU hosting available
Frankfurt-region database for EU workspaces.
Data export & deletion
Self-serve export · 30-day purge on delete.
What we collect · what we don't
| Data | Collected? | Notes |
|---|---|---|
| Agent name / ID | Yes | So you can group events. |
| Model & provider | Yes | gpt-5, claude-3.7, etc. |
| Token counts (in/out) | Yes | For spend tracking. |
| Tool call names & timestamps | Yes | For loop detection. |
| Estimated cost | Yes | Computed from public pricing. |
| Prompt text | No | Optional · off by default · scanned in-memory if enabled. |
| Model responses | No | Not collected unless you opt in. |
| End-user PII | No | Never stored. Hash or redact at the SDK. |
| Your provider API keys | No | We poll with your scoped key; raw key never stored. |
Full posture, subprocessors, retention windows, and security contact on /trust.
See how AgentWach stacks up.
AgentWach vs LangSmith
Real-time guardrails vs after-the-fact tracing.
Read AgentWach vs LangSmithAgentWach vs Langfuse
Budget enforcement vs prompt observability.
Read AgentWach vs LangfuseAgentWach vs Helicone
Loop detection + hard stops vs proxy logging.
Read AgentWach vs HeliconeAgentWach vs AgentOps
Cross-provider budgets vs agent-only telemetry.
Read AgentWach vs AgentOpsAgentWach vs Agent Watch
What "agent watch" should actually mean.
Read AgentWach vs Agent WatchAgentWach vs GitHub Copilot usage tracking
Real-time Copilot credit caps vs monthly summaries.
Read AgentWach vs GitHub Copilot usage trackingThe kinds of saves teams use agentwach for.
These are anonymised composites from customer interviews during our private beta — not attributed quotes. Named case studies will replace this section as customers go on the record.
A runaway research agent at 4am — Slack alarm fires before it burns $400. The kind of save agentwach is built for.
Agents treated like production workloads: per-agent budgets, loop guards, and replay are exactly what an SRE team expects.
Plug into a LangGraph stack in under five minutes; the injection scan catches prompts that eval suites miss.
Ready to watch your AI agents before they cost you?
Set budget limits before your agents touch production. Catch runaway costs before the invoice shows up.
No credit card · cancel anytime · free forever for solo projects