Install Hydrate
Hydrate runs on macOS, Linux, and Windows. ~30 seconds end-to-end on every platform.
1. Get the binary
macOS / Linux (Homebrew)
brew install getHydrate/hydrate/hydrate macOS / Linux (curl one-liner)
curl -fsSL gethydrate.dev/install | sh Windows (MSI)
Download Hydrate.msi (~50 MB, always the latest release) and double-click to install. Then open a new PowerShell window for step 2. Code signing for Windows Authenticode is on the roadmap: expect a SmartScreen warning on first run until then.
For verify-before-install (SHA-256 checks, ARM64 / Intel Mac tarball variants, silent MSI for CI), see the full install page.
2. Run hydrate setup
hydrate setup One command, end-to-end: mints your user id, wires hooks for every coding-agent runtime detected on this machine (Claude Code, Codex, Mistral Vibe), and offers to install the local semantic embedder (~115 MB).
3. Verify (optional)
hydrate doctor
Prints Ready. on a healthy install. If it lists red
rows, run hydrate doctor --fix to apply obvious-recovery
actions.
Want every optional knob? hydrate setup --advanced walks
the full wizard (VS Code extension, beta lock-in, BYOK keys,
enterprise registration).
4. Optional: lock $5/mo Pro forever during beta
hydrate register --edition=pro [email protected]
Pro registration is instant. hydrate register calls
our licensing service, receives an Ed25519-signed key, and activates
it locally in one CLI call. No email confirmation, no admin approval.
hydrate doctor is green, the fastest way to make Hydrate productive on a real project is the builders: one-line prompts that generate Claude Code skills, Hydrate hooks, and phased implementation plans, all wired into your local store. See /docs/builders.
What you need - and what gets better when you add more
Hydrate has two places where external services can make the product better: fact extraction (turning a session transcript into atomic architectural facts) and semantic retrieval (finding relevant facts to inject into the next session). Both have sensible defaults. Adding keys upgrades quality; you are never required to add any.
Quality tiers at a glance
| Tier | What you set up | Extraction | Retrieval | Cost |
|---|---|---|---|---|
| 0 - Minimal | HYDRATE_LLM_PROVIDER=none | Heuristic (pure Go, rule-based) | FTS5 keyword match | $0 |
| 1 - Default (what you get by just installing) | Nothing extra. claude is on PATH because you use Claude Code. | Claude Code subprocess (Haiku 4.5 via claude --print) | FTS5 keyword match | ~$0.005-$0.02 per captured session, drawn from your existing Claude Code subscription/key |
| 2 - Semantic retrieval | Run Ollama locally with an embedding model: ollama pull nomic-embed-text, then export HYDRATE_EMBED_ENDPOINT=http://localhost:11434/v1 and HYDRATE_EMBED_MODEL=nomic-embed-text | Claude Code subprocess (same as Tier 1) | Local embeddings (cosine similarity on nomic-embed-text) + FTS5 | $0 network spend (Ollama runs on your machine). One-time ~270 MB model download. |
| 3 - Cloud max-quality (what the published /benchmarks used) | export OPENAI_API_KEY=sk-... (auto-detects OpenAI for both extraction and embeddings) | OpenAI gpt-4o-mini | OpenAI text-embedding-3-small (1536 dim) + FTS5 | ~$0.0001 per fact embedded + ~$0.01-$0.05 per session extracted. PAYG on OpenAI. |
Variants: ANTHROPIC_API_KEY routes extraction straight
through Anthropic's API instead of via the claude
subprocess (slightly faster, same quality, pay-per-token). VOYAGE_API_KEY
picks Voyage's code-specialised embedding model instead of OpenAI's.
Any mix is valid - Hydrate resolves each provider independently.
Where to put the keys: don't paste them into
.zshrc. Run hydrate login openai (or
anthropic / voyage) - the CLI stores
secrets in your macOS Keychain (or a
0600 file on Linux), same pattern as
gh auth login / stripe login. Full story:
/docs/byok.
How to choose
- Just installed? You're on Tier 1. No setup needed. Works for ~80% of the value.
- Want semantic ("fuzzy") fact retrieval? Add Ollama (Tier 2). Best privacy/cost balance.
- Want the same configuration we used for the published benchmarks? Set
OPENAI_API_KEY(Tier 3). Costs pennies-per-session total. - Air-gapped / no network at all? Tier 0 -
HYDRATE_LLM_PROVIDER=none. Heuristic + FTS5. Still useful; canon injection still works.