Frequently asked.
— twelve answers, all concrete.
Does hippo work on Linux or Windows?
macOS only. Release binaries ship for arm64; Intel Macs can build from source. The capture paths use launchd, FSEvents, and macOS-specific Native Messaging plumbing — a Linux or Windows port is not on the roadmap.
Does hippo send data anywhere?
No. All inference is local — hippo defaults to oMLX (LM Studio also works) on localhost. The opt-in OpenTelemetry stack ships pointed at localhost:4317. There is no upstream telemetry endpoint.
What if I have secrets in my shell history?
Hippo redacts events before they touch the database. The default patterns catch AWS keys, GitHub PATs, generic Bearer tokens, OpenAI keys, npm tokens, and a long tail of cloud-provider patterns. See /privacy and /docs/reference/redaction.
Will hippo work with Cursor / Aider / Continue?
Yes for Cursor. Four agentic CLIs are wired in today: Claude Code, Codex, Cursor, and opencode. Aider, Continue, and others aren’t shipped sources today, but the watcher path is deliberately straightforward to extend: if a tool writes session transcripts to disk in a format hippo can parse, adding a source is mechanical. See adding a source.
How much disk space does hippo use?
About 50 MB for the daemon binaries. The SQLite database varies by usage; typical figures are 200–500 MB after a month of heavy use, dominated by the FTS5 index and embeddings. ~/.local/share/hippo/hippo.db is the only growing file.
How much memory?
The daemon idles around 30 MB. The brain calls the local inference server (oMLX or LM Studio) on demand; that server’s memory footprint depends entirely on the model you load. A typical 7B-class model on Apple Silicon uses 4–8 GB of unified memory; a 35B-A3B model uses ~20 GB.
Does the brain run all the time?
Yes, as a launchd-managed service. It’s idle most of the time, polling enrichment queues from SQLite. When work appears, it calls the local inference server (oMLX or LM Studio), writes the result, and goes back to sleep. CPU at idle is well under 1%.
What models are supported?
Any model your inference server exposes (oMLX serves anything mlx-lm supports; LM Studio loads GGUF + MLX). Currently tested with Qwen 3.6 35B-A3B; configurable in ~/.config/hippo/config.toml via the query and enrichment keys under [models]. Smaller models work; larger models work; mixing models per task works.
Can I export my data?
Yes. SQLite at ~/.local/share/hippo/hippo.db. Standard SQL. The schema is documented at /docs/reference/schema; nothing about it is proprietary.
Can I uninstall cleanly?
Yes. hippo daemon uninstall removes the LaunchAgents and binaries. The database stays at ~/.local/share/hippo/hippo.db until you remove it explicitly. The shell hook entry in ~/.zshrc is harmless when hippo isn’t running, but you can delete it if you prefer.
Is hippo open source?
Yes, MIT license. Source: github.com/stevencarpenter/hippo .
How do I contribute?
Read CONTRIBUTING.md first; then the capture-reliability rulebook at docs/capture/anti-patterns — the AP-1..AP-N anti-patterns are review blockers, not suggestions. The site itself has its own contributing guide at site/CONTRIBUTING.md .