01
Misconfiguration
- Wrong provider endpoint after a model switch
- Sandbox rules blocking legitimate tool calls
- Memory configuration pointing at deleted data sources
Service · 03
Agent Rescue.
When a harness breaks — misconfiguration, dead process, broken upgrade — we diagnose and repair it with you on the call. Secure access. Written record. Done.
Emergencies attended to in order received. For non-urgent cases, triage within the same business day.
What we fix
We've seen this before. It's probably simpler than it looks.
Most harness failures have a straightforward cause buried under layers of debug output. We know where to look.
02
Broken upgrades
- Dependencies shifted after a pip or npm update
- Critical environment variables dropped by deployment scripts
- Config format changes between major versions
03
Dead processes
- Crashes from out-of-memory or unhandled exceptions
- Model rate limits or auth token expiry halting everything
- Scheduling systems stuck in retry loops
How it works
Access that you can watch, scope, and revoke.
A rescue means trusting someone with a connection to your machine. We designed the process so that trust is verifiable at every step — and so that nothing is left open when we're done.
01
You describe it
Tell us what broke, what changed, and what you've already tried. Logs and screenshots help. Subject line "URGENT" if time matters.
02
Triage call
Thirty minutes to decide whether this is a repair, a rebuild, or something you can fix yourself with guidance. Scope agreed before work starts.
03
Secure connect
You install Tailscale and share a single node with our rescue tailnet. No credentials change hands; the key expires in 24 hours.
04
Fix together
A live session, screen shared, narration on. You see every command and can cut access at any moment.
05
Handback
Root cause, the actual fix, and what to watch for. Access is revoked when the session closes.
Your guarantees
No surprises. Nothing stays open.
- No inbound ports opened on your network, ever
- WireGuard end-to-end, peer-to-peer — no relay we control
- Ephemeral keys that expire in 24 hours, or the moment you revoke
- Credentials never leave your machine; we ask you to type them
- Full command log and session transcript, handed to you
- Scope agreement signed before any access is granted
FAQ
Questions people ask before a rescue.
How do you access my machine during a rescue?
You install Tailscale and share a single node with our rescue tailnet. No inbound ports are opened, no credentials change hands, and the key expires in 24 hours — or the moment you revoke it.
Do you hold my API keys?
No. Everything runs on hardware you own, with your keys. During a session we ask you to type credentials yourself, and nothing is stored on our side.
What do you need from me to start?
Your OS and hardware, the harness and its version, what changed last, the exact error, and what you have already tried. Logs and screenshots help.
What do I get at the end of an engagement?
A written root-cause report: what was wrong, what was changed, and what to watch for next time. Remote access is revoked when the session closes.
Can I stay on the line while you work?
Yes — and we encourage it. You see every command as it runs, can cut access at any moment, and will understand exactly what went wrong so you're not dependent on us next time.
How fast can you respond?
Emergencies are attended to in order received. For non-urgent cases, triage happens within the same business day.
Does Agent Rescue only cover harnesses you set up?
No. We work on any local AI agent setup regardless of origin — whether we configured it or someone else did. Misconfigurations, broken upgrades, dead processes, and model routing issues are all fair game.
Need help now?
Tell us what broke.
Send a short description of the system and the failure. If it's urgent, say so in the subject line and we'll prioritise triage.
Prefer to write directly? contact_us@curiositygarage.com — the address is live and monitored.
- What to include
- OS and hardware · harness and version · what changed last · the exact error · what you've tried.
- Response
- Same business day for triage. Emergencies attended to in the order received.
- Terms
- Scope and confidentiality agreed in writing before any access is granted.