How Alchemic handles your data

Which services receive what you type, upload or say, and when.

In short

• Chats are answered by AI models. Unless you keep a chat on Alchemic's own model server, what you send goes to a cloud model's maker: through OpenRouter, a model routing service, or straight to Anthropic for Claude models the platform connects to Anthropic directly. The line under the model picker says which.
• Some features send one specific piece to another service: a web search query, a database lookup, a reply read aloud.
• Your conversations, files and projects are stored on Alchemic's servers, which run on Amazon Web Services.
• This browser keeps a copy of your recent conversations while you are signed in. Signing out removes it.

Who receives what

• Cloud models: OpenRouter and the model's maker, or Anthropic

  When

  Every chat turn that runs on a cloud model (the default), deep research, and other agent work. Most cloud models are reached through OpenRouter, which passes the chat on to the model's maker, for example OpenAI, Google, xAI, Mistral, DeepSeek or Qwen. Claude models listed without a maker prefix go straight to Anthropic when the platform has its own Anthropic key. The line under the composer's model picker names every recipient for the model you picked, including the maker of the platform's routing model.

  What it receives

  Your message, the earlier messages in that conversation, text and images from files you attach, and the agents' working steps. Before an agent answers, a routing step reads your message to choose the agent; with a cloud model picked, that step runs on the platform's default model, which is usually a cloud model and can come from a different maker than the one you picked.

  Retention

  To be confirmed: the data-retention and training settings on Alchemic's OpenRouter and Anthropic accounts, and which upstream makers OpenRouter may use.

• Alchemic's own model server (models marked Local)

  When

  When you pick a model marked Local in the model picker, or keep a chat off cloud models as described below.

  What it receives

  The same as a cloud model, but it runs on Alchemic's own infrastructure and no model maker receives it. With a Local model picked, the step that picks which agent answers stays on the local server too. The local server is sometimes switched off; local models then show as unavailable.

  Agents answering on a local model can still use web search and the scientific services below, which receive search terms, not your conversation.

• Web search provider (Brave Search by default)

  When

  When an agent searches the web.

  What it receives

  The search query the agent writes from your question, sent from Alchemic's server.

• Scientific databases and literature services

  When

  When an agent looks something up or fetches a paper. The services the code can call, as of the date above:

  • Literature and papers: Crossref (including its Retraction Watch data, downloaded from GitLab), DataCite, OpenAlex, Europe PMC, arXiv, chemRxiv, Semantic Scholar, Unpaywall, CORE, HAL, OSTI, NASA's technical reports server, NSF's public access repository, USGS publications, the EPA science inventory, the Internet Archive, Zenodo, Figshare, and publishers' websites, including the Elsevier and Wiley text-and-data-mining services when the platform has keys for them.
  • Chemistry and materials: PubChem, EPA CompTox, Wikidata, the RCSB Protein Data Bank, the Crystallography Open Database, the NIST ThermoML archive and NIST Public Data Repository, SureChEMBL patent chemistry, Materials Project and MPContribs, Catalysis-Hub, and OPTIMADE databases (OQMD, Alexandria, NOMAD, NIST's JARVIS, Materials Cloud, ODBX and 2DMatPedia).
  • Life-cycle data: the US Federal LCA Commons (the USLCI database).
  • Code and models: Hugging Face, GitHub, PyPI and Anaconda (conda-forge).

  What it receives

  The search terms or identifiers the agent sends, such as a compound name, a paper's title or a DOI, from Alchemic's server. Not your conversation.

• ElevenLabs, with OpenAI as its backup

  When

  In voice mode and dictation, unless you chose Local only: while you speak, ElevenLabs Scribe transcribes what you say; when a reply is read aloud, ElevenLabs reads it in the voice you picked under Preferences, and OpenAI reads it when ElevenLabs does not answer.

  What it receives

  The recording of what you say, through Alchemic's server, and the text of each reply read aloud. Alchemic asks ElevenLabs not to keep either (its zero-retention mode); whether your plan honours that is to be confirmed. With Local only, or when Scribe is unavailable, your speech is transcribed on Alchemic's own servers instead; without either speech service, your browser's own voice reads replies, and some browser voices also run on their maker's servers.

• Your browser's speech service

  When

  Only if neither ElevenLabs Scribe nor Alchemic's own speech-to-text is available and you say yes when asked.

  What it receives

  The recording of what you say. Some browsers, Chrome among them, transcribe it on their maker's servers.

• Google or GitHub sign-in

  When

  If you choose to sign in with them.

  What it receives

  They learn that you signed in to Alchemic. Alchemic receives your name, email address and a link to your profile photo.

• Box or Google Drive

  When

  If you connect them.

  What it receives

  Alchemic reads the files and folders you choose. Opening the Drive picker loads Google's picker in your browser.

• A computing facility or lab you connect

  When

  If you connect one and run a job or an experiment on it.

  What it receives

  The job's scripts and the files it needs, or the experiment request.

• Websites you link to

  When

  When you hover over a link in a conversation, unless you chose Local only.

  What it receives

  Alchemic's server fetches the page to show a preview. The site sees Alchemic's server, not your browser.

• Amazon Web Services

  When

  Always.

  What it receives

  It hosts Alchemic's servers, database and file storage.

• Email delivery

  When

  Invitations, account activation and notices.

  What it receives

  Your email address and the message. To be confirmed: the email provider's name.

Keeping a conversation off cloud models

• In this browser: open the model picker in the composer and choose a model marked Local. Every chat from this browser then stays on Alchemic's own model server, and the browser remembers the choice, including after you sign out. The line under the picker says where the chosen model runs.
• For every chat you start: choose Local only under Default model route in your profile's preferences.

With either, a chat stops with an error while the local server is off rather than going to a cloud model. Agents may still send search terms to the services above.

Local only also covers what runs outside a chat message: the /distill-skill command, which reads a conversation's recorded steps to write a skill from them, improving a skill from your feedback, saving a conversation's lessons as a memo, a project's dream cycle, reading a reply aloud and link previews. With it chosen, these use the local server or do not run. A local model picked in the composer does not reach them, so with only that pick they can still reach a cloud model, ElevenLabs or the linked website.

Deep Research follows both too. With a local model picked, a Deep Research message runs as a normal chat on the local server instead. With Local only chosen, it runs no literature search (your question is not sent to Crossref) and its model calls stay on the local server; the report says that no sources were retrieved.

A temporary chat is not saved to your history, but it still goes to the model you picked.

In your browser

While you are signed in, this browser keeps up to 50 of your recent conversations so they open quickly. Signing out removes them, along with everything else Alchemic stored for your account; only display settings, your and your model choice stay. Other Alchemic tabs open in this browser clear the account at the same time.

Alchemic's pages load their fonts and scripts from Alchemic, not from other companies, and load no analytics or advertising trackers. The exceptions: Google's Drive picker, loaded only when you open it, and profile photos that come from Google or GitHub or were set as a web link, which load from wherever they are hosted.

Files an agent makes for you to download are not Alchemic's pages: a knowledge-graph view, for example, can load its drawing library from jsDelivr, a public code host, when you open it.

Not settled yet

• How long Alchemic keeps conversations, files and run records, and how to delete your account and its data.
• The retention and training settings of the model services above.
• Terms of service.
• A contact for privacy questions.