You can’t sail straight into the wind.
A raw model API gets you a demo. Production sits upwind: customers who can’t see each other, money collected before it’s spent, proof of every step. tack gets you there in controlled, recorded turns. You keep writing the product.
tack /tak/ · verb: to reach an upwind mark by a series of controlled turns.
You already know the moment this breaks.
You built an agent on Anthropic’s API. It works. Then you get a second customer, and a third, and you’re hand-rolling multi-tenancy in application code. Then someone asks for an invoice and you realize you’ve been eating the token cost with no meter. Then a runaway loop costs you $400 overnight.
None of that is a bug. It’s what “production” actually requires, and nobody tells you until you hit it. Ship for clients and it compounds: every client is its own tenant boundary, every retainer needs a meter you didn’t build, and every security review asks for documents you don’t have.
tack is the part that isn’t the AI.
The loop that runs the turn, who’s allowed to ask for what, what it costs, who pays, and proof of what happened. You keep writing the agent. tack is the orchestration around it you’d otherwise spend months building yourself.
One real turn, end to end.
A guest messages a hotel on WhatsApp asking to book Tuesday. Here is everything that happens before it replies: a real turn, not a diagram of one.
| 09:41:02.114 | reserve | turn_01JCM8QK · standard | hold 1.00 | |
| 09:41:02.371 | llm.turn | 2,305 tok · frontier | — | |
| 09:41:03.008 | tool.call | check_slots · 212ms · ok | — | |
| 09:41:04.101 | tool.gate | book_appointment | pending_confirmation | |
| 09:41:05.913 | confirm | POST /turns/:id/confirm | — | |
| 09:41:06.442 | usage.commit | turn_01JCM8QK | 1.00 | ⎤ txn 4189 |
| 09:41:06.442 | wallet.debit | org_01J9 | −1.00 | ⎦ txn 4189 |
| 09:41:06.501 | reply.send | "Booked. Tuesday 14:30." | — | |
| 09:47:11.083 | turn.refused | ten_09ZR · wallet_gate | 0.00 | audited |
Inside that turn: a runtime, not a wrapper.
The nine ledger rows are the outside of the turn. Inside them, an orchestration layer ran:
| Subsystem | What ran |
|---|---|
| clearance | Your backend asserts what a user earned; tack computes min(ceiling, earned) once per turn. Unknown rung falls to the ladder floor: access fails closed. |
| retrieval | Hybrid search: BM25 and vector legs fused after per-leg access control, never before. The reserved never class is structurally unretrievable. |
| the loop | A durable turn queue, transcript memory, and pre-emit guards running inside the stream, not after it. |
| tools | Typed contracts, confirmation gates, per-tool timeouts that resolve to error envelopes. Repeated failure raises tool_degraded: failure is routed, not accidental. |
| judgment | Guardrails are semantic judges, not regex lists, screening output before it is emitted. |
| the books | The usage row and the wallet debit commit in one transaction. Both or neither. |
That layer is the expensive part. Not the model call: the clearance, retrieval, judgment and failure routing around it — the maturity a team buys with thousands of engineering hours. On tack it is already running in production, and you spend those hours on your product instead.
Join the waitlist
This isn’t a vibes problem. It’s a measured one.
Ask any founder running an AI product and they have a margin story. Measured across hundreds of companies, it looks like this:
| Measured | Finding | Source |
|---|---|---|
| 84% | of companies report 6%+ gross margin erosion from AI infrastructure costs; 26% report 16% or worse. | Benchmarkit / Mavvrik, 2025 State of AI Cost Management (372 enterprises), via PR Newswire and Forbes |
| 80% | of enterprises miss their AI cost forecasts by more than 25%. Only 15% land within 10%. | Same Benchmarkit / Mavvrik survey, 372 enterprises, 2025 |
| ~60% | average gross margin for AI-native “Shooting Stars”, the fastest-growing as low as 25%, against 70%+ for traditional SaaS. | Bessemer Venture Partners, State of AI 2025 |
| +36% | one-year jump in average monthly AI spend ($63K → $85.5K), while only half of teams say they can track the ROI on it. | CloudZero, State of AI Costs 2025 (500 respondents) |
tack’s wallet exists for exactly this: hard caps and per-turn metering before the money leaves, not a forecast you’re hoping holds.
Join the waitlist
What skipping this costs you.
And it doesn’t cost anything to find out. Metered has no monthly fee: you pay only on turns you’re already billing your own customers for.
Join the waitlistRequest the security pack
Money, or momentum. Pick the card that’s yours.
Ship your agent this week instead of your billing layer for six months.
Every turn rolls up per tenant. Price your product per turn: the spread is your margin.
It’s a real API.
Two calls between a signup and their AI answering.
For teams that charge money.
Create a tenant. Submit a turn. Everything else was already running.
Pre-launch. The prices on the home page are the real prices; keys go out to the waitlist first. Blocked by a security review? Say so: the threat model, DPA and sub-processor list go out by reply. We only use your email to notify you — see the privacy policy.