Nightly early access. Nightly builds only for now.
Hexbot

Providers and billing

Hexbot runs the agent. Model providers run the models. You bring provider credentials and the provider bills your account for usage. Hexbot does not include model tokens or pay those charges.

Add a provider

Open Settings, Providers and choose a provider.

  • Subscription providers, including ChatGPT or Codex, SuperGrok, and Nous Portal, sign you in through your browser. Follow the provider’s sign-in page and enter the displayed code if it asks for one.
  • API key providers take a key from your account with them. Hexbot saves the key for the daemon. A saved key does not guarantee that your account can use every model in the list; send a short test message after choosing one.
  • Local or externally managed providers may require endpoint or credentials configuration on the daemon computer. Follow the setup message shown for that provider.

Credentials belong to the daemon deployment, not to one bot. Every bot on that daemon may use any configured provider. Setup walks you through the first provider and then lets you add more before it asks for a default.

Default and fallback models

Provider settings also hold two deployment-wide choices. The default model is pre-filled whenever you create a bot. The optional fallback is the model every bot switches to when its own provider is down or rate limited.

Treat provider keys like passwords. Do not paste them into a chat, a bot persona, or a project file. Hexbot stores credentials in the daemon’s private configuration. Replacing a provider key updates open sections before their next provider request.

Choose a model

Each bot has a provider and model. The picker lists only providers you have set up, with a short curated list on top. For OpenAI Codex, OpenRouter and local servers it shows the models the provider offers right now; other providers use the catalog that ships with Hexbot. A section can also carry a temporary model override.

Each bot also has a reasoning level in its Model settings. The list shows only the levels the chosen model supports, for example Low, High and Max for GLM-5.3 Flash. Higher levels think longer before answering and use more tokens. The default is Medium, or the next level up the model supports. Choose it when you create a bot, so the bot’s first section starts at that level.

A new model or level applies from the next message in every section still on the previous one. The first reply after a switch rereads the conversation without the provider’s prompt cache, so it costs more than usual.

LM Studio lists the models available at its local /v1/models endpoint when you open its model picker. Start LM Studio’s server before choosing a model. If discovery fails, the daemon’s current configured model stays in the list.

Local and custom servers

Start the model server on the daemon computer, or at an address that computer can reach. localhost means the daemon’s computer, even when your client app runs elsewhere.

For an OpenAI-compatible custom server, set the model fields in the daemon’s config.yaml. Merge them into the existing configuration rather than replacing the whole file. For example, an Ollama server can use:

model:
  provider: custom
  base_url: http://127.0.0.1:11434/v1
  default: your-loaded-model

Replace your-loaded-model with an actual model name on that server. Refresh provider settings and start a new section after configuration changes. For an authenticated endpoint, configure its credentials on the daemon too. Local models need the capabilities your task uses, such as tool calling or image input.

Tell the daemon the model’s context window when it cannot find out itself. When a section’s context fills, the daemon summarises its oldest part, starting at 75% of the window or later. The reserve defaults to 16,384 tokens and the recent context kept defaults to 20,000 tokens; both are capped at a quarter of the model’s window. A model the daemon does not know is assumed to have 32,768 tokens, so a larger model would be summarised far earlier than it needs to be.

For Ollama, set context_length under model. The daemon sends it to the custom endpoint as num_ctx and uses that window for every model there, including fallback models:

model:
  provider: custom
  base_url: http://127.0.0.1:11434/v1
  default: your-loaded-model
  context_length: 32768

For another server that manages its own window, use model_overrides instead of context_length:

model:
  provider: custom
  base_url: https://your-server.example/v1
  default: your-loaded-model
model_overrides:
  custom:
    your-loaded-model:
      context_window: 131072
      max_output_tokens: 8192

model_overrides.<provider>.<model> also accepts supports_reasoning and supports_vision. A _default entry under a provider applies to every model of that provider the daemon has no metadata for. On the custom endpoint, model.context_length takes precedence over these window overrides. Project files are never read for this: a .pi directory in the workspace does not change how a section is summarised.

LM Studio’s default endpoint is http://127.0.0.1:1234/v1. Start its server and load a model before selecting it in Hexbot. Use the LM Studio provider for its model discovery.

Model names, prices, context limits, and availability can change. Check the provider’s own pricing page before using an unfamiliar model or enabling long unattended tasks.

Model credentials and connectors

A model provider generates replies. Web search, cloud browser, image generation, and premium voice are separate connectors with their own credentials and per-bot switches. If a bot can reply but cannot use one of those tools, check its Connectors settings instead of replacing the model key.

Understand charges

Providers commonly charge for input and output tokens. Tool results, conversation history, memory, and long files can increase input size. A bot that calls many tools may make several model requests for one visible answer.

Hexbot does not add a fee to provider requests. Your provider dashboard is the source of truth for invoices and quotas.

Rooms, private teammate exchanges, delegated subtasks, scheduled jobs, and dreaming can all make additional model requests. Settings, Usage reports token counts and estimated costs; estimates may differ from provider invoices. Room limits and a daily budget for each bot cap what rooms spend.

If a provider fails

  • Confirm the selected model is available to your account and its quota has not been exhausted.
  • For subscription providers, sign in again if the provider reports expired or revoked access.
  • For a local server, confirm it is running and reachable from the daemon computer.
  • Check the fallback choice when a provider is down or rate limited. It needs its own valid credentials.
  • Check tool-specific errors in Connectors when the model itself still replies.

Remove access

Clear a provider key, or sign out of a subscription provider, in settings to stop new requests through that provider, including requests from open sections. Bots configured for it will need another model before they can reply.