What LLMs Can I Use With Handle?

There is no model picker in Handle, and no API key to supply. We choose the model, we change that choice as better ones ship, and the honest answer to “which model am I on?” is “whichever one is currently best for the work in front of your Handler”.

---

You Don’t Pick a Model

Every other integration in Handle is something you connect. Models are not: they’re part of the platform, the way the database behind your Handler is.

That’s deliberate. Picking a model well means knowing which ones are strongest at long reasoning this month, which are fast enough for a reply someone is waiting on, which handle a 200-page PDF, and what each costs per task. That answer changes every few weeks. It’s our job to track it, not yours.

---

How Handle Chooses

The choice is made per step of the work, not per Handler and not per account. A single task routinely uses more than one model:

• The hard reasoning — working out what a messy email is actually asking for, or whether an invoice matches a purchase order — goes to the strongest model available for it.
• The high-volume, well-defined steps — classifying, extracting fields, summarising a thread — go to lighter, faster, cheaper models that are entirely capable of them.
• Anything a person is waiting on in Slack is weighted towards speed, because a correct answer that arrives four minutes later is a worse answer.

So the trade-off being managed is between intelligence, speed and cost, on every step. Sending everything to the largest available model would be slower and more expensive without being more accurate at the work that doesn’t need it.

---

Kept Current

New frontier models ship every few weeks, and the ranking between them changes with each release. We evaluate them as they land and move Handlers onto the ones that win, on our side.

For you, that means:

• No migration. Nothing to upgrade, no version to bump, no Handler to reconfigure.
• No pinned versions to track. There’s no model setting in Handle that quietly becomes last year’s default because nobody revisited it.
• No setting to get wrong. The most expensive model isn’t the best one for every step, and this is not a decision you can lose by making it badly.

We don’t publish a pinned list of model names and versions, because the list would be out of date within the month — and because which model handled which step is not something you need to hold in your head to trust the result.

---

Why Handle Doesn’t Support BYOK

Handle does not support Bring Your Own Key. There’s nowhere to paste an API key and no way to attach your own provider account. The reason is architectural rather than commercial.

Handle runs on several providers at once, and reaches some models by more than one route. Which provider serves a given step is ours to choose and ours to change — for a better model on that step, and to keep more than one way of reaching a model available to us. Routing is a thing we move, not a thing we settle.

A key you supply is the wrong shape for that:

• A key belongs to one provider. It authenticates against one vendor’s API, and a single task routinely crosses more than one vendor as it moves between steps. The key cannot follow the work.
• It would pin you to one vendor. Your Handlers’ availability and capacity would become that vendor’s availability and capacity, with none of the alternate routes we keep open behind them.
• It carries its own limits. Your account’s rate limits and quotas would become your Handlers’ limits, and a burst of work would meet them at exactly the moment it mattered.
• It would need constant attention. Every routing change we make — a new model, a new provider, a region shift — would become a key you had to authorise, or something that broke without warning.

Model access comes with Handle instead: there’s no provider account to hold and no inference invoice to reconcile.

---

What This Means in Practice

• There’s nothing to set up, and nothing to keep up to date.
• Your Handlers get better over time without you doing anything, as the models underneath them improve.
• If you have a compliance requirement about where inference runs or which providers are involved, that’s a conversation to have with us directly rather than a setting to change.

See also: Supported Integrations for the platforms your Handlers can connect to, and the security model for what a model is and isn’t trusted with — including why credentials never reach one.