Tell is a Rust-based gateway between your product and every LLM provider. It caches, routes, redacts, and audits, at sixty thousand requests per second, with sub-millisecond overhead.
No SDK rewrite. No vendor lock-in. No quiet model swaps that hit your customers before they hit your dashboard.
POST /v1/chat/completions · json_mode · ctx 32k · region us
When this matters
Tell is for teams whose product talks to a model on every meaningful click, and whose ops, finance, and legal teams are starting to notice.
One SDK, one bill, one failure mode. The model that beat yours on price last week sits on the other side of a rewrite you keep deferring.
The same prompt is generating the same answer for the tenth time today. Two teams are calling GPT-5 for tasks 4o-mini would solve. There is no budget by team, no cap by project.
A model gets deprecated. A fingerprint flips overnight. A region quietly throttles. You find out the same way your customers do: through the support inbox.
PII rides along in the prompts. Secrets land in the logs. There is no audit trail that survives a regulator, and no redaction layer that survives a code review.
How it works
Tell sits on the wire. Every prompt your product makes, every response a model returns, passes through the same place, and gets the same treatment.
Tell speaks the OpenAI API. Swap one base URL in your existing client; keys, types, and streaming all keep working. No SDK rewrite, no new abstractions to learn.
A constraint solver filters providers by region, context length, cost ceiling, and JSON mode. A multi-objective bandit then scores the survivors on cost, latency, and quality.
A semantic cache check answers similar prompts in under a millisecond. PII and secrets are redacted at the edge. The chosen provider streams the response back in coherent chunks.
Prompt, params, route decision, tool traces, fingerprints, confidence map, all attached to one trace. Sealed originals; redacted day-to-day logs; audit-ready exports on demand.
What ships
Three things drop into place the day Tell goes live in front of your stack.
Constraints first, scoring second. Fewer wasted GPT-5 calls on tasks 4o-mini would handle; fewer truncated 4o-mini calls on prompts that actually needed the bigger context.
Vector lookups on incoming prompts. Forty to sixty percent of traffic stops short of a provider; the answer was already on the shelf, ten characters of paraphrase away.
Prompts, params, route decisions, fingerprints, redactions, all in one trace. The forensics bundle attaches itself to the incident, not to the engineer who happens to be on call.
“summarise this 12k-token transcript”
“classify ticket sentiment”
↳ served from semantic cache
“extract invoice fields → JSON”
↳ json repair · 1 fixup
“draft customer email re: refund”
↳ PII redacted · 2 spans
POST /v1/chat/completions · json_mode · ctx 32k · region us
Why Tell
The same problems show up in every LLM stack. Solving them once, on the wire, beats solving them four times in four apps.
01
It lives on the wire. Every request, every retry, every failover passes through the same place, so policy, cost, and audit are written once, not re-implemented per app.
02
Model fingerprints watched, behavioural contracts enforced, shadow traffic compared. When a provider quietly swaps a checkpoint, you find out before your customers do.
What Tell isn't
The category is full of half-fits. Here's what makes Tell structurally different.
It routes on cost, latency, and quality. It repairs JSON output to schema. It hedges tail latency across providers and cancels on first success. A reverse proxy doesn't.
Your code keeps using the OpenAI client. Tell is the URL it points at. No new types, no rewrites, no second SDK to lock yourself into.
Request, response, prompt diff, route decision, cost, confidence: same trace. One pane, not three integrations and a Looker dashboard.
Common questions
No. Tell speaks the OpenAI API. Point your existing OpenAI client at Tell's URL: keys, types, streaming, tool calls all keep working. The migration is a config change, not a project.
Hosted, per-region, with a global control plane and per-region data planes. Or BYOC inside your VPC, where the data plane stays in your residency zone and the control plane stays out.
Smart fallback. Tell holds the chosen model, hedges to a backup if tail latency spikes, and routes around outages without changing the request your client sent. The first you'll hear about it is the dashboard.
PII and secrets are redacted at the edge before egress. Day-to-day logs are redacted; sealed originals sit in immutable storage with break-glass controls and per-tenant KMS keys.
Yes. Self-hosted endpoints (vLLM, TGI, in-house Triton, anything that speaks OpenAI or a known schema) join the routing pool alongside hosted providers; same constraints, same scoring, same audit.
Pricing
Cached calls are free. Routed calls are billed at provider cost plus a thin margin. The incentive sits on the right side.
Plan 01
Free · 90 days
For closed-beta partners
Drop Tell in front of one product surface. All features unlocked, weekly check-ins, and pricing locked in for the first year.
Plan 02
From $499/mo
For shipping LLM teams
Hosted, multi-region, full feature set. Usage-based billing on completions that actually leave the gateway; cached calls are free.
Plan 03
Custom
BYOC data plane
Single-tenant deployment in your VPC. BYOK encryption, dedicated SRE, hard multi-tenancy across business units.
§ 03 · Engagement intake
01 / Start a brief
We staff small and senior, scope by phase, and end on a written deliverable. We don’t sell decks or hours.
If we’re not the right team for the job, we say so on the first call. The bar is production, not pitch.
02 / Where to find us
Montreal, Quebec
Satellite office