tamag0
Documentation

Model providers

Your models, your choice: plug in Claude, Codex, Ollama, or any OpenAI-compatible endpoint — or go fully local. The model layer is a choice, not a constraint.

Interchangeable runtimes

Runtime Models Where inference runs
Claude (default) Anthropic models via Claude Code Anthropic API
Gemini Enterprise Agent Platform (Vertex AI) Gemini and Claude A Google Cloud project you designate
Codex OpenAI models via Codex OpenAI API
Ollama Any Ollama model Any endpoint you configure — local, LAN, self-hosted, or Ollama Cloud
OpenAI-compatible vLLM, Together.ai, and similar Any endpoint you configure — local or remote

From the agent's point of view, runtimes are interchangeable: same memory, same tools (files, shell, web, integrations), same skills, same permission gating, same collaboration features. Switching provider doesn't reset who the agent is.

Reuse the subscriptions you already pay for

No separate API billing required. tamag0 signs in to the accounts you already have:

  • Claude — sign in with your Claude account (OAuth): usage runs on your Claude subscription. An Anthropic API key works too, if you prefer metered billing. The desktop uses Claude Code's standard sign-in flow, so reauthentication is the same account flow as signing in from Claude Code itself. If the installed Claude Code is too old to provide that flow, tamag0 asks you to run claude update before retrying instead of attempting an unsupported login.
  • Codex — sign in with ChatGPT: usage runs on your ChatGPT Pro or Plus subscription quota, not OpenAI API credits.

At every launch, the desktop updates Claude Code and Codex CLI in the background. If an update is unavailable or fails, tamag0 still starts normally.

Codex models

On Auto, Codex turns use three GPT-6 models with a Low, Medium, or High reasoning effort (Luna runs at Medium or High):

Model Suited to Manual efforts Default effort
GPT-6 Luna Short, targeted tasks at low cost Low to Max High
GPT-6.2 Sol Everyday development and code review Low to Ultra High
GPT-6.2 Astra The most demanding professional work Low to Ultra Medium

A new conversation starts on GPT-6.2 Sol at High, and so does a conversation with several participants, unless someone picks a model. After that, the agent re-evaluates the model and effort its work needs at each response. The X-High, Max, and Ultra efforts, and the GPT-5.6 Sol, Terra, and Luna models, are manual choices only.

GPT-6.2 Sol replaced GPT-6.1 Sol at the same price per token, and GPT-6.2 Astra replaced GPT-6 Astra at a lower one: conversations on the earlier models move to their GPT-6.2 successor automatically. Both need Codex CLI 0.169.1 or newer; with an older CLI, tamag0 asks you to update it from Settings → LLM Providers → ChatGPT instead of starting the turn.

If your team already pays for Claude and ChatGPT, your agents run on those plans from day one — zero additional inference budget to negotiate before trying tamag0.

Gemini Enterprise Agent Platform (Vertex AI)

Use one Google connection for Gemini, Claude, or both, depending on what your project offers. Inference runs in the region you select and is billed to your Google Cloud project. You don't need a separate Claude account. The Gemini and Claude setup below requires Desktop 1.88.0 or later.

  • Where: Settings → LLM Providers → Vertex AI. The Context menu at the top of the tab chooses whose access you configure: your home workspace (shown by its name) or one of the clients where your agents are on assignment. Each context has its own Google connection and its own configuration.
  • Connect: select Sign in with Google. tamag0 opens the Google sign-in in your browser through the Google Cloud CLI (gcloud), which must be installed on the computer; if it is missing, tamag0 says so instead of starting a sign-in. You don't run gcloud auth application-default login yourself.
  • What you enter: the Google Cloud project ID (the ID, not the display name), the region, and the models you want to offer — see Choose the models below. Then select Save configuration, and Verify access to check that the connected account can use the project.
  • Use it: add Vertex AI to a runtime priority — Settings → General for a context, or Settings → Agent → Per-agent runtime override for one agent — or pick it in the composer like any other provider.
  • One connection per context: the home connection and each client's connection are kept apart. Connecting, reconnecting or disconnecting one never changes another, and none of them uses the Google account already signed in on the computer, another gcloud configuration, or an environment variable.
  • Renewal and reconnect: access renews automatically while Google's authorization is valid. If it was revoked or expired for good, the provider shows Reconnection required and a Reconnect Google button for that context, and turns stop with that message; tamag0 does not switch to another provider.
  • On this computer: every Google connection stays on the computer where you made it, in a private folder managed by tamag0; it is not sent to tamag0's servers or synced. A client's project, region and models follow your account to your other computers; the home workspace's stay on this one. On another computer, the same context asks you to connect again rather than borrowing the Google account already signed in there. Change Google account replaces that context's connection, and Disconnect deletes it; neither touches another context.

Choose your Gemini and Claude models

You don't need to type a model name: fetch the models available to this context, then choose which ones to offer. The list lives under Models in the configuration, next to the project and the region. It can include both Gemini and Claude; a Google connection is shared by both families.

  1. Check access. On first setup, enter the project ID and region. tamag0 automatically checks access to models in both families. When you reopen a saved configuration, select Show all models to check access; use Refresh to repeat the check. Results distinguish models that are available, not enabled for the project, blocked by a data-sharing requirement, over quota, or unavailable in the selected region. Being listed in the catalog does not mean you have access.
  2. Choose models. On first setup, recommended Gemini models that pass the access check are selected automatically. You can change that selection and choose up to 32 models. A selected Claude model that is no longer in Google's catalog can still appear in the list. If the catalog is unavailable, you can Add a model by its ID.
  3. Give each role to a ticked model. The configuration needs at least one ticked model and has two roles, each chosen among the ticked models: - Default model — what a conversation runs when nobody picked a model. - Small background tasks — the short tasks the agent runs outside the answer. This role is for Claude models; Gemini uses the selected model for its turn. Same as the default model is the default choice when the default is Claude. Unticking a model that holds a role moves the role to a remaining compatible model.
  4. Save configuration. Nothing is offered in conversations until you save. A configuration saved before this list existed shows its model and its Haiku model already ticked.

Ticking is not enabling. A model must be available to your project and region. The access check reports the reason when Google refuses a model, so you can distinguish project access, data-sharing, quota and regional availability issues before choosing it.

Region. The model list and access checks use the region you set. A model not served in that region cannot be used there; select a supported region or choose a model that is available in your region.

If the fetch or access check fails, tamag0 says why and offers Try again. Google may deny the account, block a region endpoint through organization policy, report a model as unavailable, or return a quota/data-sharing restriction; resolve the displayed cause with your Google Cloud administrator where needed.

In a conversation. Open the execution menu of the composer and choose Vertex AI: a sub-menu lists the ticked models by family, with the default marked default. Pick a model to run the conversation on it; with no pick, it runs the default model. For Gemini models that support it, open Effort to choose a reasoning level. When you switch models, an effort that the new model doesn't support returns to Auto. In a client context, the sub-menu offers the models ticked for that client.

Runtime priority. A runtime priority can name a specific Vertex model, not only the provider. An individual agent can use a different Vertex model from the global priority. Where automatic fallback is available, you can list one Vertex model and then another. Existing provider-only entries continue to mean the configured default model.

If this computer isn't connected to Google, a client context keeps its Vertex AI entry but it reads Connect in Settings and can't be chosen. At Home, Vertex AI is offered only once it is configured and this computer is connected to Google; until then the entry is absent from the menu.

Agents already set up for Vertex

The older per-agent checkbox (Claude → Via Google Cloud (Vertex AI)) is replaced by the provider. After the update, each agent that had it keeps its exact project, region and models as its own Vertex AI configuration for the home workspace — two agents with different settings keep theirs apart — and its own runtime priority becomes Vertex AI only. Your global priority, other agents and assignments are unchanged; the old setting is not moved to a client. If that agent has no Google connection yet, Vertex AI stays selected and asks you to connect — it does not fall back to your Claude account. An entry that cannot be read is shown for repair rather than deleted. If one of your custom providers is already named vertex, the migration waits and the agent's turns stop until you recreate that provider under another name; nothing is renamed automatically.

Fully local when you want it

With Ollama or a self-hosted OpenAI-compatible endpoint, your agents' inference never leaves your infrastructure — for companies that need complete control. Local endpoints can be keyless; cloud endpoints keep their keys in the OS keychain. This covers the agent's model only: tamag0's own processing of company memory (search indexing, summaries, memory extraction, nightly cycle) runs at DeepInfra and Together AI in the United States (see Architecture). Both processing providers apply zero data retention to covered requests; see Security for what each policy covers.

Per-agent choice and automatic fallback

  • Each agent can have its own runtime and model — a coding agent on Claude, a triage agent on a local model.
  • A runtime priority order tells Auto which configured providers to try. Before execution, tamag0 skips providers that cannot be used locally, such as a missing or signed-out CLI, a disabled runtime, or an incomplete custom-provider configuration. These preflight skips are not runtime attempts and do not appear in the attempt history.
  • Runtime fallback is narrower than preflight availability. During an Auto turn, tamag0 continues to the next entry in the priority order only when it recognizes a capacity or rate limit. Examples include recognized Claude or Codex quota/session exhaustion, temporary Claude overload (HTTP 529), a model capacity limit on Vertex AI, and an HTTP 429 from Ollama or an OpenAI-compatible provider. It does not infer exhaustion from ordinary assistant prose.
  • Authentication, network, timeout, other HTTP 5xx, context-window, cancellation, and unknown failures stop the turn. They do not start another provider. In particular, a network failure is never treated as automatic fallback.
  • When a capacity failure happens after visible output or a tool action, tamag0 keeps the ordinary conversation history, persists visible partial output when there is any, and starts the successor with a short internal continuation instruction. It never mechanically replays a completed tool. A durable switch notice records the provider change.
  • Capacity failures pause the affected provider for that agent, or just the affected model on Vertex AI. This cooldown survives desktop restart. An explicit provider reset time wins; otherwise tamag0 uses a valid Retry-After value, and allows another attempt on a later Auto turn after one minute when neither is available. Expired cooldowns are removed automatically. A successful attempt, including a forced provider/model selection or Retry, immediately clears that agent’s cooldown for the provider or Vertex model that succeeded, so Auto can select it again in priority order.
  • On an assignment that uses the client's access, there is no fallback at all: a failed turn stops with its error, whatever the cause. See LLM access on an assignment.
  • Automatic fallback applies only while the composer is on Auto. An explicit provider or model pin stops on failure. Retry bypasses that provider's cooldown for one attempt only and never starts a fallback cascade; Switch to Auto and retry returns control to the configured priority order.
  • Your global priority and each agent's priority order follow your account to another device after restart. Gemini Enterprise Agent Platform (Vertex AI) Google connections, and the home workspace's Vertex AI project settings, are the exception: they stay on the computer where you made them. Local choices keep working offline; if an offline change could not be saved, the account's last saved choice is restored at the next online start.
  • Provider availability is resolved dynamically from what is actually installed and authenticated — the app never forces a default you don't have.
  • Adding an Ollama or OpenAI-compatible provider lets you fetch its model list and pick the models to add. On a provider you already saved, fetching reuses the saved key — you don't retype it, and it is never sent to the interface. A company can also declare custom providers and models shared across its agents.

Right-sized execution

Model and effort are selected per task: routine work runs on lighter, faster settings; deep architecture or debugging work gets the strongest model and higher reasoning effort. Cost stays proportional to task complexity — and both onboarding installs (Claude and Codex) are set up out of the box so the choice stays open.

  • Performance — execution efficiency and usage visibility
  • Security — key storage and isolation