The incident report for your agent's tool calls — before the deletes happen.

agentguard sits between your agent and its MCP servers. It classifies every call (read / write / spend), halts loops, caps writes, emails and dollars per run, fakes destructive writes in dry-run and shows you the diff, and writes a hash-chained audit log with daily Merkle roots on a public page. The OSS proxy does all of that locally; hosted adds the dashboard, alerts, the ledger and the "what would this run have destroyed" report.

Sign in with GitHubWatch the live demo
1 · Create a proxy (20 s)
Paste your MCP server URL. You get a per-customer endpoint and a key.
2 · Connect (30 s)
npx @agentwares/agentguard connect agk_…
Writes .mcp.json for Claude Code / Cursor, or paste the URL into any MCP client.
3 · Watch the first run
Calls stream into the dashboard. Dry-run fakes the writes and shows what would have changed. Flip to enforce when it looks right.

Bands, not a meter

Soft alert at 80%. Hard stop at 150% with a one-click upgrade — never a surprise bill.
Trial
$0/mo
  • 1,000 tool calls/mo
  • hosted proxy, dashboard
  • no card
Starter
$29/mo
  • 100,000 tool calls/mo
  • 30-day audit log
Pro
$99/mo
  • 1,000,000 tool calls/mo
  • 90-day audit log
  • least privilege
  • spend ledger
  • chaos mode
  • drift alarm
  • pii audit
Team
$299/mo
  • 10,000,000 tool calls/mo
  • 365-day audit log
  • least privilege
  • spend ledger
  • chaos mode
  • drift alarm
  • pii audit
  • policy sync
  • approvals
  • compliance export
  • erasure

Machine-readable: /pricing.json. Agents: /llms.txt, MCP at /api/mcp. Proxy endpoints look like https://agentwares-agentguard-proxy.vercel.app/<proxyId>/mcp.

What the agent sees when it is stopped
{"code":"LOOP_DETECTED",
 "cause":"crm_get_contact was called 4 times with identical arguments within the last 30 calls (max 3)",
 "fix":"you are looping. Do not repeat this call; change the arguments, use a different tool, or stop and report what you tried",
 "retryable":false}
What you see
An incident-shaped run report: calls, writes, faked writes with the intended mutation, blocked calls with their code, spend by rail, the chain verification and the day's Merkle root. Alerts by email, Slack or Discord when a run is halted or a band hits 80%.