Skip to main content

AI engines

An AI engine is what powers an agent's thinking. OtoDock doesn't ship a model — you bring your own subscriptions and keys, and choose the engine that fits each job. There are three, and you can mix them freely across agents and even switch per chat.

The two that do the real work​

OtoDock runs your agents on the two most capable AI systems available — Claude Code and Codex. These aren't chatbots that return text; they're full agents that read and write files, run commands, browse the web, use tools, plan multi-step work, and keep going until the job is done.

They were built for software engineering — one of the hardest things to automate well — and that same depth makes them excellent at complex work of every kind: research, analysis, operations, content, support, and more. You connect your own subscription and your agents run on it.

Claude Code​

Anthropic's coding agent, running on Claude models. Connect it with a Claude Pro or Max subscription or an Anthropic API key. It's the default starting point for most agents and supports the full feature set — streaming reasoning, plan mode, to-do tracking, sub-agents, and live tool use.

Codex​

OpenAI's coding agent, running on GPT models. Connect it with a ChatGPT subscription or an OpenAI API key. Codex is a strong alternative or complement to Claude Code — set up some agents on one, some on the other, or give a single agent both and pick per chat. It is also the engine that runs local models as full agents.

Run agents in a live terminal

Both Claude Code and Codex can run as an interactive terminal session right in the dashboard — you watch (and steer) the agent's native interface live, just like running it on your own machine, with all of OtoDock's files, tools, and security around it. See Live terminal.

The direct engine​

The third engine calls AI provider APIs directly. It's lighter-weight and lower-latency, which makes it the right choice for phone calls and voice conversations, where every fraction of a second counts. It also works in chats and tasks when you want a quick, simple response rather than a full agent.

The direct engine is multi-provider: it runs on Anthropic, OpenAI, Groq, or your own local models, and it also names your chats. An admin can choose the model that writes the titles, or turn titles off, under Setup → System Settings. Agents on it can read and write the agent's files, use the platform's tools, and search the web through the provider's own API on Claude and OpenAI models. It doesn't do plans, to-do lists, sub-agents, or the live terminal; for day-to-day agent work, Claude Code and Codex are the ones to reach for.

Local and self-hosted models​

Running your own models? On a self-hosted install an admin adds a local endpoint once, under Setup → AI Engines → Local models, and it serves both Codex and the direct engine:

  • Ollama — give it the server's OpenAI-compatible address, http://<host>:11434/v1.
  • Any OpenAI-compatible endpoint — llama.cpp, LM Studio, vLLM, and similar, by URL.

An API key is optional, for servers that require one. Discover lists the models the server offers so you can add them in one go. On a Docker install, address the server by its LAN IP or host.docker.internal, never localhost (see Local network access).

What to expect from a local model:

  • Codex or the direct engine. Claude Code has no local-model path. A Codex agent can run a local model even on a remote machine; that machine then needs to reach the server itself.
  • Tools on Codex depend on the server. Codex needs a server that speaks OpenAI's Responses API. Ollama 0.34 and newer, vLLM, and LiteLLM carry the agent's tools through; on Ollama, Codex loads them on demand, so a large toolset no longer bloats the first prompt. llama.cpp and LM Studio drop the tools, and the agent is told so. A local model that needs tools on one of those belongs on the direct engine.
  • Thinking shows in the chat, and the agent's effort setting is honoured: Low turns a local model's thinking off, anything higher leaves it on.

Choosing the engine and model​

Each agent has a default engine and model, set in its configuration. You can:

  • Give an agent more than one engine and choose between them per conversation.
  • Switch the model from the dropdown in any chat — no restart needed. While a conversation's session is live, the dropdown offers the active engine's models; once it has ended, the agent's other engines appear too — pick one and the chat continues on the new engine, carrying its history across.
  • Choose the thinking effort that fits each agent (from quick to maximum) in its settings — so it digs deeper on hard problems or stays quick on simpler ones.
  • Pin a single scheduled task to a different model or engine than the agent's default — a cheap default with one demanding daily job on a stronger model, or the reverse. See Tasks.

Each engine's default, chosen by tier​

Every model has a capability tier: frontier, strong, balanced or fast. Model lists sort their models by tier and mark each one with dots, and hovering a model shows its tier and what it's good at.

An agent without a pinned model uses its engine's strongest model below the frontier tier. Frontier models are always there to pick, but never the default, so choosing the most capable one is always deliberate. An agent set to Auto shows which model it's running, and if an admin turns that model off, the agent falls back to the next one down the tiers, never up.

Admins can add their own custom models and local endpoints, and give each one its tier and a one-line note on what it's good at, under Setup → AI Engines.

OtoDock keeps the model lists current as Anthropic and OpenAI release new models. When a new version replaces an older model, OtoDock swaps it in everywhere: the model lists offer the new version, and the agents, chats and scheduled tasks set to the old model move to it the next time they start, with nothing for you to change.

Bring your own, or share a pool​

Engines are connected by the people who own them, and your own chats and tasks always run on your own connections first. You can connect several accounts per provider — each appears with its email, and a per-connection switch lets you rest one from your own work without disconnecting it. Admins can also contribute connections to a shared pool for work that belongs to an agent rather than a person — agent-scoped scheduled tasks and meetings, and shared agents — so it has an engine to run on. (Your own user-scoped tasks and meetings still use your own connection.) Both users and admins can cap subscription use — see Usage & Limits.

Prefer pay-as-you-go? Connect your own Anthropic or OpenAI API key under User Settings → AI Engines (one key per engine) and cap what it spends per week or month.

→ Full walkthrough: Connecting your AI.

Next steps​