open-context open-context
October 21, 2026 · 6 min read

Why open-context Uses Ollama Instead of a Hosted LLM API

open-context uses Ollama instead of a hosted LLM API because its AI preference-analysis step takes your entire exported chat history as input — the most personal text you've typed to another assistant — and the only way to avoid trusting a third party's retention policy with that input is to not send it to one. Ollama runs the model on infrastructure you choose, local or remote, so the single network call open-context's analysis can make goes to an address you configured, not a vendor's servers. Skip AI analysis entirely with --skip-preferences and the whole pipeline makes no network calls at all.

Key takeaways

  • open-context's OllamaPreferenceAnalyzer only ever calls the Ollama host you configure — there is no built-in code path to a hosted LLM API.
  • --skip-preferences produces a full markdown conversion with zero network calls; AI analysis is opt-in, never required.
  • Hosted APIs like OpenAI's exclude API data from model training by default, but still retain prompts and responses for up to 30 days of abuse monitoring — a window local inference skips entirely.
  • Samsung's 2023 internal ban on generative AI tools, triggered after engineers pasted source code into ChatGPT, shows this isn't a hypothetical risk for sensitive text.
  • open-context's own FAQ leaves the door open to a future Claude API option, but Ollama stays the default because it keeps analysis local by design, not by configuration.

What the AI analysis step actually sends somewhere

Most of open-context's pipeline never touches the network. Parsing your ChatGPT conversations.json, normalizing it into the shared schema, and writing markdown files all happen in-process on your machine. The one stage that calls out at all is the optional AI analysis step, implemented in src/analyzers/ollama-preferences.ts as OllamaPreferenceAnalyzer. It defaults to model gpt-oss:20b against http://localhost:11434, checks that Ollama is reachable and the model is pulled, then sends your conversation text — capped at roughly 100,000 tokens, about 400,000 characters, regardless of export size — inside a prompt asking for a preferences or memory summary.

That's the entire outbound surface. There's no telemetry, no analytics beacon, and no secondary call to validate a license or fetch a config. If checkOllamaAvailable() can't reach the host, the analyzer throws immediately rather than silently falling back to something else.

Why not just call a hosted LLM API?

A hosted API would be simpler to integrate than managing a local model. The tradeoff is that it adds a retention and training-policy dependency to a tool whose entire input is your private conversation history. OpenAI's own developer docs state that data sent to its API has not been used for model training by default since March 2023 unless a customer opts in — but the same docs note that abuse-monitoring logs, which can include prompts and responses, are retained for up to 30 days by default, with a zero-data-retention option available only to approved customers who apply for it. (Source: OpenAI's "Your data" guide)

That's a reasonable policy as hosted APIs go. But it still means trusting a third party's retention window, abuse-monitoring exceptions, and approval process for anything stricter. Running Ollama locally removes that dependency instead of just tightening it — there's no retention window to reason about when the request never leaves your network.

Real-world incidents show the risk isn't hypothetical

In 2023, Samsung banned employees from using ChatGPT and similar generative AI tools on company devices after engineers uploaded internal source code to the chatbot to check for errors, and others used it to summarize confidential meeting notes. Samsung's stated concern was that these services store input on external servers, where sensitive data is hard to retrieve or delete once submitted. (Source: TheWrap) A tool like open-context exists specifically to process years of chat history — the exact category of data that incident was about — which is why its default analysis backend doesn't ask you to make that same tradeoff.

How to run open-context with zero network calls

If you want a hard guarantee rather than a policy to trust, skip AI analysis entirely:

npm start -- convert path/to/chatgpt-export.zip --skip-preferences

This produces the full conversation markdown and a basic, non-AI preferences and memory summary generated from conversation statistics alone (message counts, date ranges, title keywords) — no model call, no network request. If you do want the richer AI-written summaries, point --ollama-host at infrastructure you control, including a remote GPU box on your own network:

npm start -- convert export.zip --ollama-host http://gpu-server:11434

Running the Docker image works the same way. Ollama on your host machine is automatically reachable at host.docker.internal:11434, or set OLLAMA_HOST to point the container at a different server entirely.

Will open-context ever support a hosted API?

Possibly. The project's own FAQ already addresses this directly: Claude API support "may be added later," but it isn't available today, and Ollama is free and runs locally in the meantime. If a hosted option does arrive, the expectation set by the rest of the project is that it would be an explicit opt-in alongside Ollama — the same way AI analysis itself is already opt-in via --skip-preferences — rather than a replacement for the local-first default.

Convert your chat export with analysis that never leaves your machine.

Get started with open-context →

FAQ

Does open-context ever send my chat history to OpenAI, Anthropic, or another hosted API?

No. Today, open-context's only AI analysis backend is Ollama, and it only talks to the Ollama host you configure (locally by default, or a remote server you control via --ollama-host). It has no built-in code path to a hosted LLM API.

What exactly gets sent to Ollama during analysis?

Your conversation text, capped at roughly 100,000 tokens (about 400,000 characters) per run, plus a prompt template asking for a preferences or memory summary. Nothing else about your machine or account is transmitted.

Can I guarantee zero network calls during conversion?

Yes. Running the CLI with --skip-preferences produces the full markdown conversion of your conversations without calling Ollama or anything else — the pipeline falls back to a basic, non-AI preferences and memory summary generated entirely offline.

Is Ollama itself sending my data back to a company?

No. Ollama is open source software that runs model inference on the machine or server you install it on. Once a model is pulled locally, generating a response does not require a call back to ollama.com or any other third party.

Will open-context ever support a hosted API like Claude's instead of Ollama?

Not yet. The project's own FAQ notes that Claude API support may be added later, but Ollama remains the default specifically because it keeps the analysis step local — any hosted option would need to be explicitly opted into, not the default path.

For a deeper look at how this fits open-context's broader architecture, see Self-Hosted vs. Cloud AI Memory: Where Should Yours Live.