Skip to contentdocs.Growther.si
Download C5

Model providers

A provider is a source of SI models. C5 needs at least one before it can do anything.

You are not locked into a single company. You can connect several and use them for different jobs.

Adding a provider #

  1. Go to Settings.
  2. Open the Integrations tab.
  3. Pick the provider you want.
  4. Paste in your API key.
  5. Click Save.

C5 checks the key immediately and tells you if it does not work.

OpenRouter is set up differently. Its card in Integrations has no key field: C5 reads its key from your .env file. The card shows whether the connection works, and the account's credits and usage. Its routing settings have a tab of their own; see OpenRouter routing settings.

Which models a provider lists. A cloud provider lists the models its own listing named the last time C5 asked it. Before the first answer, after a failed one, or without a key, the model pickers fall back to C5's built-in list. MS Azure lists only the deployment you configured. For your agents, OpenRouter lists the models you picked under Model Selection on its Config tab. The classifier's model list is the exception: it offers every model OpenRouter itself lists (see What the classifier's questions carry).

Where keys come from #

An API key is a long password that lets a program use a service on your behalf. You get one from the provider's website, usually in a section called API keys or Developers.

Warning Treat an API key like a password. Anyone who has it can spend money on your account. Never paste one into a chat message or an email.

Your keys are stored encrypted on your own computer and are only ever sent to the provider they belong to. See Local encryption.

Using many providers at once #

Connecting more than one is a good idea:

  • Different strengths. Some models are better at some things.
  • Different prices. Cheap models for easy work, strong models for hard work.
  • A backup. If one provider goes down, C5 can use another.

See Choosing models for how to match jobs to models.

Running models on your own computer #

You can run models locally with no outside service at all. This means:

  • No cost per use
  • Nothing leaves your machine
  • It keeps working with no internet

The trade-off is that local models are usually less capable than the best paid ones, and they need a reasonably powerful computer.

To connect one, choose the local option under Integrations and point C5 at where your local model is running. C5 checks the connection and lists the models it finds.

Gateways #

A gateway is a service that gives you many models through a single key. If you use one, connect it like any other provider and C5 will list everything it offers.

This is a simple way to try lots of models without signing up everywhere.

OpenRouter routing settings #

While OpenRouter is turned on in Settings › Integrations, Settings has a tab for its routing settings. The tab is labelled Config, with OpenRouter's icon; hover it to read OpenRouter Config. (Each provider and gateway you turn on can add a Config tab of its own, so look for the icon.) The settings are grouped as Auto-Router & Models, Complexity Routing, Provider Preferences, Hardware & Data and Plugins, and Raw JSON shows the whole file, openrouter_config.json. Nothing applies until you press Save, which needs permission to edit settings.

Price caps #

Max Prompt Price and Max Completion Price, under Provider Preferences, are in US dollars per million tokens. They are hard limits: OpenRouter uses only providers at or below them, and a request fails when no provider is that cheap. 0 means free providers only. A negative or non-numeric entry is saved as no limit.

A stored 0 waits for you. An earlier version of this tab showed a stored 0 as blank, so a 0 could be saved without anyone seeing it. Until you confirm such a 0, C5 leaves it out of chat and task requests, and Provider Preferences says so: Not applied to chats and tasks: Max Prompt Price is 0, which allows free providers only. … Choose Keep 0 as the limit or Clear the 0, then Save. Saving another setting does not confirm it. Questions the classifier sends to an OpenRouter model apply a stored 0 either way, so they use free providers only.

Other settings #

  • Model Selection: the models to use, in order. A request carries at most three: the agent's own model first, then the next ones in this order. The models you pick here are also the OpenRouter models your agents can choose from. They are not the classifier's choice: its list offers the models OpenRouter lists that it could choose, and choosing one there never adds it here or changes the order a chat or task tries.
  • Allow Fallbacks: whether OpenRouter may use a backup provider for the same model when the first is unavailable (true by default). It is not about fallback models; those are the Model Selection list.
  • Sort Strategy: Price, Throughput or Latency, or -- (Balanced). Global Sort Partition ranks the providers of every model in the list together, and needs a sort strategy: while Balanced it is off and nothing is sent for it.
  • Data Collection: deny routes only to providers that do not collect your data. It is not Zero Data Retention, a stricter OpenRouter filter you set in Raw JSON as "zdr": true.
  • Response Healing: OpenRouter repairs malformed JSON in a reply, only for replies that are not streamed and were asked for as JSON.
  • Middle-out compression has no control on the tab: set "middle_out_compression": true in Raw JSON, and chat and task requests carry OpenRouter's context compression.

A reply is asked for as JSON only where the work needs one, for every agent, whether or not this tab was ever saved. Up to v329, once the tab had been saved, a request from an agent whose name contains planner was always asked for JSON, which broke the plans and routing answers C5 reads line by line; such plans now work as with every other provider.

Provider lists and plugins #

  • Provider Order, Allowed Providers and Ignored Providers take OpenRouter provider slugs — who serves a model, such as deepinfra or together — not who made it. The tab shows them; set them in Raw JSON as "order", "only" and "ignore".
  • Plugins: the switches under Plugins turn a plugin on or off by its name ("web", "response_healing", "context-compression"). To give a plugin options, write it in Raw JSON as an entry with an id, such as {"id": "web", "max_results": 3}; to keep one off, give it "enabled": false. A save from the tab keeps such an entry with its options. Its switch shows what requests carry: on, unless the entry says "enabled": false. Flipping that switch changes only enabled, and keeps the options.

When a save is refused #

If a value cannot be used, nothing is saved, and the message names the first problem, for example OpenRouter settings were not saved: sort must be one of price, throughput, latency, exacto (got "fastest"). While Raw JSON holds something that does not parse, the tab says so and keeps the last valid settings.

If C5 cannot read these settings #

Your price caps and provider limits travel with every chat and task request to OpenRouter. If C5 cannot read them at the moment of a request — the database did not answer, or what is stored is not a settings object — it never sends that request without them:

  • It sends the limits it last read since C5 started, so a passing fault changes nothing you would notice.
  • If it has not read them once since starting, it refuses the request instead: C5 cannot read your OpenRouter routing limits (OpenRouter's Config tab in Settings), so it did not send this request: it never sends one without your price caps and provider limits. Check them in OpenRouter's Config tab in Settings. The request is not retried on its own.

(That is the Config tab with OpenRouter's icon, the same place the classifier's notes name.) C5 writes this to its log once per start, not once per request. If the refusal persists, open the Config tab: if the tab cannot show your settings either, run growther doctor.

Fleet › Assignment never erases these settings either. Giving agents an OpenRouter model there saves it with the rest of OpenRouter's settings, so when C5 cannot read them first, it changes nothing and says so: Could not set M for OpenRouter: C5 could not read its current settings, so it changed nothing. The same holds for the local model pool. Try again once the settings can be read.

What the classifier's questions carry #

The classifier's model list for OpenRouter (see A model from your providers) offers the models OpenRouter lists, read from OpenRouter's public list of models, with your key, through your proxy if you set one. C5 reads it only where a classifier model from OpenRouter could be chosen — Models allowed from at Anywhere, OpenRouter turned on, and your organisation's policy allowing it — and offers only the models that policy's allowedModels allows. C5 keeps the list for ten minutes. If it cannot read it, the list shows only your Model Selection models and says so: C5 could not read the models OpenRouter lists, so this shows only those under Model Selection on its Config tab. Choosing one here never changes that list. Choosing or testing a classifier model never changes these settings, and never changes which models a chat or task tries.

A classifier model chosen from OpenRouter is asked with your data collection, Zero Data Retention and distillable-text settings, the provider order and the allowed and ignored providers, quantizations, the sort, the price caps (a stored 0 included), and Allow Fallbacks when it is false. It is never sent the fallback models or the plugins, and C5 asks OpenRouter to use only providers that support every part of the question. If C5 cannot read these settings, it asks OpenRouter nothing they might rule out.

If OpenRouter requests fail after an update #

Up to C5 v326, the price caps and the sort were sent under names OpenRouter does not read, so they had no effect. From v327 they work, so a cap below the price of the model a request is routed to now makes that request fail. Check Max Prompt Price and Max Completion Price on the Config tab, and confirm or clear a stored 0 there.

When a provider fails #

The usual causes:

  • The key is wrong. Copy it again — trailing spaces are a classic.
  • The key expired or was revoked. Make a new one.
  • You are out of credit. Check your account with the provider.
  • The provider is down. Check their status page.
  • You hit a rate limit. C5 slows down and retries by itself.

To re-check a provider's key, open Settings → Integrations and save it again. C5 tests the key as you save and tells you straight away if it does not work.

A coordinator that keeps failing is tried last. When 3 of its last 5 answers come back as an error or empty, or 3 of its last 5 attempts get a server error, a refused connection or a dead gateway, C5 moves it to the end of the order for a while: a minute at first, doubling each time it fails again, up to 15 minutes. Saving the provider again does not end that; a good answer, the time running out, or restarting C5 does. Hover a chip for why and for how long:

  • Skipped — repeated failed generations: its answers came back empty or broken.
  • Tried last — repeated gateway errors, in amber beside its status: its gateway answered with server errors, refused the connection, or was down. Its status check can still answer, so the card may read Online at the same time. It is still tried, after every other coordinator.

When both apply, only Skipped is shown. See Choosing models.

Where to go next #

Edit this page on GitHub ↗Last updated Nov 20, 2026