Day one: your API. Day five: read-only tools and the eval in your CI, or refunded. Credited in full.
Your official connector for Claude, ChatGPT and Cursor.
Every write waits for your user's approval, in code you own.
Inspect the connectorWatch a write wait.

Your decision
comes first.
Ask the agent for a transfer.
The connector holds it here.
Current balances
Teller · CADInteractive preview with demo data. No real money.
The gated connector
First three companies with a named reference and short case page. CAD 6,500 without one or afterwards.
Balance only when your signed sandbox acceptance test passes.
Nightly eval, API change alerts and a written fix date. Cancel any month.
One build at a time, evenings and weekends.
Agree the checks.
Then build.
Trial fixtures: blocked writes, denied or expired approvals, duplicates, failing API replay.
Isolated in your repo. Trial API access stays read-only.
- Tools marked read or write.
- Every write held for approval.
- Your exact-match threshold on 80 cases.
- Replay without duplicate writes.
- Your engineer reviews the write path.
On your infrastructure.
- Code
- Your repo, day one.
- Access
- Your sandbox and test tenant.
- Credentials
- I never hold production credentials.
- Approval
- Your queue. Your users decide.
- Audit
- Redacted calls and decisions in your store.
A reply within two business hours on weekdays.
Replies only. Fixes planned that evening; longer fixes get a written date.
On failure: a safe error, writes held in your queue.
The eval method,
in my own product.
Duskline, my live evening planner, uses a language model behind an MCP server.
Inspect Duskline ↗Exact matches over six rounds. Invalid outputs: four to zero.
AgentOps Gate and Capstan: my approval and replay work.
I'm Affan. I build connectors. Check the code above. Write to me about your API.
affanyasir9@gmail.com