Local-first by design.
Last updated 2026-07-27 · Apache-2.0
- SuperBased is built so the data it captures about your AI coding sessions stays on your machine by default.
- The default install is telemetry-free - no analytics beacon, no required account, and nothing cloud-based runs unless you opt in.
- There is one opt-in exception: the hosted Cloud Intelligence service (see the Privacy Policy), which only ever sees what you explicitly sign in for and consent to upload.
This page describes exactly what runs, where data lives, and what changes if you turn on Teams mode or sign in to the optional hosted Cloud Intelligence service.
1. What runs locally
The observer binary is a single Go executable that runs three things on localhost:
| Component | Listens on | What it does |
|---|---|---|
| Watcher | no port - reads files | Tails the session-log directories of supported AI coding tools (e.g. ~/.claude, ~/.codex, ~/.cursor) as new JSONL events are appended. |
| Local proxy | 127.0.0.1:8820 |
Opt-in. Sits between an AI client and the LLM provider to capture token counts you can't get from the rollout log alone. Disable any time. |
| Dashboard + MCP server | 127.0.0.1:8081 |
The web UI and the stdio MCP endpoint your AI agent can call to query its own history. |
Everything SuperBased stores lives in one SQLite file at ~/.observer/observer.db. Delete the file and SuperBased's state is gone.
2. What does NOT leave your machine
- The contents of your prompts and responses.
- The files your AI agent read or edited.
- Your project paths, git remotes, and branch names.
- Token counts, costs, and timing.
- The SuperBased DB itself.
The default install makes zero outbound network calls beyond the ones the AI client makes anyway to its own LLM provider (which SuperBased just observes when the proxy is enabled).
3. Teams mode - opt-in, hash-by-default
SuperBased supports an optional Teams mode where multiple developers push their session metadata to a self-hosted org server. The wire shape is designed around two privacy invariants:
Hash by default
Every content-bearing column (paths, commands, project roots, git remotes) ships only as a SHA-256 hash. The hash lets the server aggregate “how many developers touched this file” without ever seeing the path.
Raw content is node-side opt-in
Raw values also ship when the node's own config sets either of two keys:
# ~/.observer/config.toml - on the developer's own machine [org_client.share] full_content = true # or admin_managed = true
admin_managed is for admin-provisioned fleets, where collection is configured at the source. Neither has a remote toggle - both are node-local, and the node can override them. Even so, raw_tool_input, raw_tool_output, preceding_reasoning and error_message are never read by the seam.
One audited SQL seam
The single SQL seam that builds wire rows is internal/store/orgpush.go::SelectUnpushedSince. The privacy-strip happens row-by-row in Go before sending. Any new content-bearing column has to be gated there and added to the regression sentinel in tests/invariant/privacy_test.go.
4. Source & audit
The full source is published under the Apache License 2.0. You can read every line that decides what gets observed, what gets persisted, and what (under Teams mode) gets put on the wire.
- Source on GitHub
- Every binary release ships with a CycloneDX SBOM and an SLSA-signed provenance attestation.
- Releases are signed; verify with
slsa-verifier ≥ v2.7.0.
5. Reporting a vulnerability
Email [email protected] with reproduction steps. We acknowledge within 72 hours.
6. Operator entity
SuperBased is published by Gaja AI Private Limited. Correspondence: [email protected].