Connect a model

abstract doesn’t come with a model of its own. Connect at least one in Settings, then pick the model in the chat box. You can connect several and switch at any time.

Your ChatGPT plan

If you have ChatGPT Plus or Pro, abstract can run on your plan. No API key needed.

1. Open Settings and click Continue with ChatGPT.
2. A new tab opens at OpenAI. Sign in and allow abstract to use your plan.
3. The tab closes itself and Settings shows ChatGPT plan · active with your email.

The first time, a short notice explains that abstract’s requests count toward your plan’s limits. You can see and cap what abstract uses at chatgpt.com/settings/usage. The Manage usage link in Settings and next to the model picker goes there.

Good to know:

• Plan usage must be allowed. If you declined it during sign-in, Settings shows plan usage off; click Continue with ChatGPT again and allow it.
• ChatGPT plans don’t include embeddings, so library search falls back to keywords unless OpenAI, Google, or Ollama is also connected (see Embeddings).
• Usage limits. When the limit is reached, the chat shows a Usage limit reached card with Manage usage and Switch model. Everything done so far is saved.
• Sign out revokes abstract’s access at OpenAI. Use a different ChatGPT account signs in with another account.
• A sign-in must be finished within 10 minutes, and before you restart abstract.

API keys

Open Settings → + Add provider, choose a provider, and paste the key:

Provider	Key from
Anthropic	console.anthropic.com
OpenAI	platform.openai.com
Google	aistudio.google.com
OpenRouter	openrouter.ai (one key for many vendors)

Keys are stored in ~/.abstract/config.json on your computer. You can also set them as environment variables, which take precedence:

export ANTHROPIC_API_KEY=…
export OPENAI_API_KEY=…
export GOOGLE_GENERATIVE_AI_API_KEY=…
export OPENROUTER_API_KEY=…

To change or remove a key, click manage next to the provider.

A local model with Ollama

Ollama runs open models on your own machine, so document content never goes to a cloud provider.

1. Install Ollama, start it, and pull a model, for example ollama pull <model>.
2. In Settings → + Add provider → Ollama (local), enter the server address, usually http://localhost:11434. (Or set OLLAMA_HOST, which takes precedence.)
3. Your pulled models appear in the model picker.

abstract’s agent relies heavily on tool calling, so choose a model that supports tools; results depend a lot on the model’s size and quality.

Choosing the model

Click the model name under the chat box. Models are grouped by provider, and only connected providers appear. Each list shows recent models only (the newest version of each model family), so it stays short.

One model does everything: planning and writing, checking claims, reading scanned pages, and sub-agents all use the model you pick. When a ChatGPT model is selected, the chat box shows Using ChatGPT plan · Manage usage.

If you haven’t picked one, abstract uses a default from the first connected provider, in this order: ChatGPT plan, Anthropic (newest Sonnet), OpenAI (newest GPT), Google (newest Gemini Flash), OpenRouter, Ollama.

Note

Whichever model checks a claim, a supported verdict still needs a verbatim quote that the code finds in the source. A weaker model tends to produce more partial verdicts, not false greens.

Embeddings

Library search combines keyword search with semantic search. Semantic search needs an embedding model, which abstract picks automatically; it’s never shown in the picker:

1. OpenAI text-embedding-3-small, if OpenAI is connected,
2. else Google gemini-embedding-001,
3. else Ollama nomic-embed-text.

With none of these connected (for example, only a ChatGPT plan), search is keyword-only. Everything else works the same.