OpenScience

Privacy

Last updated 16 September 2026. Applies to OpenScience and the Synthetic Sciences services it can connect to.

OpenScience is a local application. Your projects, sessions, credentials, and results are stored on your machine. Prompts are sent to the model provider and scientific services you choose to call, and signed-in session traces are shared with Synthetic Sciences as described below. If you sign in to Synthetic Sciences, we receive what an account and a Wallet require, we log the routing and cost of managed requests so we can bill you, and we log session traces to improve the agent. Trace sharing is on by default, including sessions using your own provider connections. Saved opt-outs are preserved. You can turn it off in General settings or from your account at any time, and everything we hold about you can be deleted. This page describes each of those flows in the order data moves.

What this page covers

The OpenScience software (the command line tool, the browser workspace, and the desktop app), the openscience.sh website, and the Synthetic Sciences account services that OpenScience can connect to: sign-in, the Wallet, Ace managed models, managed research search, and trace sharing. It does not cover the model providers, scientific databases, or other services you connect. Each of those is governed by its own policy.

What stays on your machine

Your working copy is local. The agent runs as a server bound to your loopback address with a Host and Origin allowlist, and there is no remote mode. The data root defaults to ~/.openscience and can be moved. It holds your sessions, messages, and tool outputs as files; your artifacts, figures, and the provenance graph that links each result to the code and sources that produced it; your provider keys, connector credentials, and OAuth tokens, encrypted at rest and never uploaded to us; and a per-inference manifest of hashes for the system prompt, instructions, and tool schemas, so a result can be traced to the exact configuration that produced it. The agent reads and writes your project files only inside the workspace and the paths you grant.

Shell commands, Python and R kernels, and local compute jobs run inside macOS Seatbelt or Linux bubblewrap when containment is on, with network egress denied. Known credential patterns are redacted from tool output before it is written into a session. The full picture is in the security documentation.

What leaves your machine

Model and tool requests go directly from your machine to the service you invoked unless you chose a managed route. Separately, while trace sharing is enabled, the client sends a redacted record of session activity to Synthetic Sciences.

  • Your model provider receives prompts, attached files, tool outputs, and the system prompt on every model call, under that provider's policy.
  • Scientific databases and connectors receive the query the agent issues, for example a UniProt accession or a literature search, when a tool calls them, under each service's terms.
  • MCP servers, compute backends, and publishing targets you configure receive whatever the tool you approved needs, including any files it uploads, when you approve the action.
  • GitHub receives a request for release metadata when OpenScience checks for updates.
  • Synthetic Sciences receives account identity, Wallet activity, managed requests, and session traces, only while you are signed in, under this page.

Your own provider keys and provider sign-in credentials are not uploaded to Synthetic Sciences. Activity using those connections, local models, Python and R, SSH, or your own Modal account may appear in a shared session trace, including prompts and tool inputs and outputs.

What Synthetic Sciences collects

Signing in creates a device key bound to your account and, if you choose one, an organization workspace. From then on we hold four kinds of record.

Account. Your email, sign-in identity, workspaces, and the devices you have approved. You can revoke any device from the dashboard.

Wallet. Purchases, holds, settlements, and promotional credits, recorded in micro-dollars. Card details are handled by our payment processor and never reach our servers.

Managed requests. For Ace and managed research search we record the route, provider, model, token counts, and cost of each request, and the funding context it ran under. This content-free usage record is what your usage charts and invoices are built from. The request body passes through to the provider and is kept only as part of a session trace, under the rules below.

Shared credentials. If your workspace shares provider credentials, OpenScience refreshes an encrypted, short-lived overlay of them. Your local keys are never uploaded, and a local key always wins over a shared one.

Session traces

A trace is the full trajectory of a session: your prompts, the model's reasoning and answers, every tool call with its inputs and outputs, and the route and model that served each step. Traces are how we find where the agent goes wrong and how we build the benchmarks on the home page. Trace sharing is on by default while you are signed in, and you can turn it off at any time.

Usage records retain the token counts and costs reported by the provider, with reasoning and cached tokens kept as separate reported details rather than added again to the total. A missing amount is marked unavailable; the client does not substitute a generated count or a catalog-price estimate. These records are diagnostic evidence, not an independent verification of a provider invoice. Managed billing continues to use the gateway's settlement records.

While sharing is on, five rules apply.

  • Batches are uploaded compressed and authenticated with your device key, and each delivery carries an idempotent id, so a retry can never duplicate an event.
  • The client redacts known credentials and secret-shaped fields before writing its upload queue, and the ingest path applies its own redaction before storage. Redaction does not remove all personal or research information from a trace.
  • Raw trace events are kept for 30 days and then purged. A content-free usage projection of route, model, token counts, and cost outlives that window.
  • Access to raw traces is limited to security and super-admin roles with a fresh second factor, and every access is audited.
  • Each stored event carries the hash of its canonical payload and a receipt id. The client checks the delivery id and acknowledged event ids before removing a queued record; an unverified response leaves the record queued for retry.

Upload queues and individual records have size limits. Oversized content is marked as truncated; binary attachments are represented by metadata, and logging does not open additional files to upload them. Delivery failures and rejected records are shown in General settings. An interrupted process may leave an incomplete trace.

Turning sharing off

Trace sharing has a device switch and versioned account preferences. In General settings, turn off Share session traces to stop this device's uploads and discard its queued and rejected records. This does not delete your local conversations or records already received by the service. The client requires a supported account disclosure version before uploading and preserves saved opt-outs.

Three switches control it. Analytics covers content-free usage: routes, models, token counts, and timings. Research content covers the prompts, outputs, and tool payloads themselves; complete trace sharing requires both this and analytics, and the server rejects a request that sets one without the other. User-owned routes decides whether sessions that ran on your own keys, a ChatGPT sign-in, a local model, or a custom endpoint are included at all, independently of the other two.

Change any of them from your Synthetic Sciences account. Account opt-outs are enforced at ingestion and checked before each upload. Turning off user-owned routes excludes those traces while allowing managed routes under your remaining preferences. Signing out stops uploads from the device.

Deletion and retention

You can delete your account data from the dashboard. Deletion removes raw traces, ingest receipts, the usage projection, and aggregates in one transaction, and records a durable opt-out so a client that was offline at the time cannot re-upload later. The deletion request and the consent history are kept as compliance records; they contain no research content. Independently of deletion, raw trace events expire after 30 days, and Wallet and billing records are kept for as long as the law requires us to keep financial records.

This website

openscience.sh sets no cookies and runs no third-party analytics. Fonts are served from this domain. The only cross-origin request the site makes is to GitHub's public API, from your browser, to show the repository's star count in the footer. The site is hosted on Vercel, whose edge network keeps standard request logs.

Security

Local credentials are encrypted at rest. The local server accepts connections only from your machine. Ace requests carry their funding context, and the client rejects any response whose scope does not match, so a personal device key can never spend an organization's funds. Our full threat model is in SECURITY.md. Please report vulnerabilities through the GitHub security advisory form.

Changes and contact

This page is versioned in the OpenScience repository, so every change is public and dated. Material changes to trace collection also bump the consent version, which means you are asked again before anything new is collected. Questions or requests about your data: privacy@syntheticsciences.ai.