teaflask
Backed byY Combinator

Make your product AI-native by tomorrow.

Introducing Teaflask

Your users don’t want your interface. They want what’s behind it. Every menu, every form, every help article is a translation layer between intent and outcome, charging a little interface tax on every click.

Teaflask removes it. It gives your product an AI assistant grown from your own code: it knows how the product works, meets users inside it, and carries their intent through to done.

Unlimited capability

Anything your users can do in your product, Teaflask can do for them.

Submit

Click anything

Vendor
Marriott Berlin

Fill forms

Auto-approve

Change settings

invoice.pdf68%

Upload files

March
910111213141516171819202122

Schedule things

unpaid invoices⌘K

Search records

TripTotal
Berlin1,142.80
Lisbon318.40
Madrid907.05

Read data

Attach receiptFile report

Complete tasks

POST/v1/payouts

Call APIs

Issue the refund?

CancelRefund

Confirm actions

Most chatbots stop at answering questions. Teaflask can execute complex multi-step workflows. If your product allows it, teaflask can do it.

Docs. It answers from knowledge grown out of your own code, re-verified every time that code moves.

Actions. It makes only the API calls scanned out of your product, each one left on Allow, Ask, or Deny.

Workflows. It follows your own playbooks, written once, calling those actions by name.

Settings. You choose the model and the memory, and whether the turn runs on your key, your credits, or the user's own plan.

Batteries included

Built for the enterprise.

The runtime, the guardrails, and the UI an assistant needs inside a real product, already built and already connected.

Durable execution

Crashes, restarts, two-day approvals — the run picks up mid-step.

Sub-agents

It splits long work across background sub-agents and merges their reports.

Real-time streaming

Streamed token by token. A dropped connection resumes mid-sentence.

Auto-scaling

Capacity steps ahead of the spike; nothing sits in a queue.

Sandboxed execution

A run reaches only the tools and hosts admitted to it.

Security

Tenant isolation enforced in the data layer, not in application code.

Assistant UI

Composer, threads, attachments and model choice — shipped, not assembled.

Conversation analytics

Every action it took, every approval, and what the turn cost.

Models and keys

Any model, on your own key or the user’s own AI plan.

Governance plane

Held actions wait in a queue until a person approves them.

Provenance engine

Every claim names the code behind it, and re-checks when it moves.

SimulationComing soon

Replay saved conversations against a changed workflow before your users see it.

Live in a day

Run your product on autopilot.