Cove.AI
New buildRescueManaged
Start a build →
Rescue · audit-first · consent-gated

Something’s wrong
with your agent.

The bill is bigger than it should be. Or something quietly stopped working. Or the whole thing just feels off. We’ve seen it before, we know where to look, and we fix it without touching anything you don’t approve.

Real audit, recently
$250/day
in surprise API charges
One file in the agent’s memory had quietly grown too big to read. Every conversation re-sent the whole thing. The dashboard looked fine. The bill didn’t.
Sound familiar?

If two or three of these match,
we should talk.

01

Your bill is bigger than it should be

Cost leaks the operator can’t see from the dashboard.

  • API spend keeps climbing for no obvious reason$250/day observed

    A common one: the agent’s memory file has quietly grown past what the runtime can actually read. Every conversation re-sends a giant payload that gets silently trimmed before the agent ever sees it. Bigger bill, smaller brain. We’ve seen this drive $250/day in unintended Anthropic charges on a single instance.

  • Background jobs you forgot were running

    Old automations still firing into the most expensive model around the clock, long after whatever they were doing stopped mattering.

02

It’s broken and nobody noticed

Things that fail without raising an alarm.

  • A core feature stopped working weeks ago

    Auth tokens expired. A scheduled sync hasn’t actually run. A channel hasn’t actually reconnected. There’s no error — just a quiet absence of the thing it used to do.

  • The docs say one thing, reality is another

    Your config claims a sync runs every hour. The system has no record of running it in six weeks. The agent answers questions based on stale data and sounds confident doing it.

03

The agent is confidently wrong

Memory and context have rotted.

  • It still cares about projects you cancelled months ago

    Terminated work, expired deadlines, dead initiatives — all still loaded into the agent’s working memory and influencing every reply.

  • It treats old urgencies as live

    A deadline from March is still flagged urgent in May. The agent prioritizes ghosts and asks you about things you finished or abandoned.

04

Your fleet has drifted apart

For people running more than one instance.

  • Hard-won rules on one machine, missing on the other

    Security guardrails, channel rules, behavior fixes — you remember setting them up, but they’re only in place on one instance.

  • Old experiments still hanging around

    Sub-agents and integrations from earlier ideas, never cleaned up, slowly drifting further out of sync each week.

05

Something feels off but you can’t name it

Soft symptoms that usually point at real problems.

  • The chat UI keeps disconnecting

    Reconnect loops, dropped sessions, websockets falling over. Usually a small misconfiguration that’s also burning agent runs every time it cycles.

  • You assumed a problem was a bug. It isn’t.

    Some things you’re fighting are actually security features doing their job. We tell you when to stop trying to fix them.

+ 20 more signals we audit silently.Show the technical list →

For the engineers reading this — these are the actual signatures we sweep on a rescue audit. If any of them mean something to you, you already know why we look:

  • MEMORY.md size & per-file injection limits
  • KeepAlive sidecars and cron entries
  • Model tier routing for vision vs. text
  • Fallback chains when MLX/Ollama dies
  • auth-profiles.json state
  • Bonjour name collisions
  • Stuck session ghosts and orphaned .jsonl
  • Plugin runtime deps and upgrade debris
  • Doctrine asymmetry across hosts
  • SecretRef vs. inline secrets
  • gateway.bind scope
  • Cross-context messaging guardrails

Findings are written up in plain English. The technical detail is available if you want it; not required reading if you don’t.

How it works

Three steps. Nothing changes without consent.

  1. 01

    A short call

    You tell us what feels off. We tell you what an audit costs.

  2. 02

    We look. You approve.

    Read-only audit. You see every finding before anything is touched.

  3. 03

    We fix what you greenlight

    Then hand back a healthier instance and a written report.

Stop the bleeding.

A read-only look can’t make things worse.

It can show you exactly what’s wrong, and what it’ll cost to fix.