enzyme

enzyme [-p|--vault PATH] [--workspace NAME] [-v] command [options]

enzyme is the Enzyme CLI: a local-first compile step for Markdown knowledge bases (Obsidian vaults, meeting notes, agent memory). It reads how notes are already organized (folders, tags, links, and dates), writes a short editable `.enzyme` program, and builds a local index so a coding agent can retrieve related passages and cite them.The invoked binary is enzyme (Cargo `[[bin]]` in crate `enzyme-cli`). Workspace package version 0.13.1. Search runs on the machine. Catalyst generation (questions about what you decided, changed, or left open) may call Enzyme's hosted service or a local GGUF model.A typical Markdown setup is compile (propose settings), init (build the index), doctor (report what was indexed), then petri (map) and catalyze (search). refresh picks up later notes. install writes agent skills and instruction markers for claude, codex, hermes, or openclaw.Named workspaces live under `~/.enzyme/workspaces/` plus the workspace name (override the home with ENZYME-HOME). --workspace / --collection selects one and cannot be combined with -p / --vault. Without either, the vault is the current directory; legacy vault data is `vault/.enzyme/` (database `enzyme.db`, log `enzyme.log`).

-p, --vault PATH

Vault directory. Defaults to the current directory. Global. Conflicts with --workspace

Named workspace declared under the Enzyme home. NAME is a single filesystem component (no slashes). Global. Conflicts with --vault

Raise log verbosity. Repeatable. Default is warnings only; once is info, twice is debug, three times is trace. Global. enzyme compile -v prints the generated program

Print the package version and exit

Print clap help for the CLI or a subcommand

~/.enzyme

Enzyme home when ENZYME-HOME is unset. Holds `config.toml`, `models/`, workspace programs in `configs/`, named workspace data in `workspaces/`, and apply journals under `configs/.enzyme-apply/`

Legacy per-vault data: `enzyme.db`, `enzyme.log`, `refresh.lock`, `workspace.lock`. Named workspaces store the same files under `~/.enzyme/workspaces/`

Readable settings compiled by enzyme compile. Declares `workspace`, `source markdown` / `source sqlite`, and `learn questions from` readings with a profile (`operational`, `decisiontrace`, `resonancetrace`, `reflective`, `relational`, `preferenceevidence`, or default `tensiontrace`)

Used only with --llm env. OPENAI-BASE-URL `https://openrouter.ai/api/v1` takes precedence for Markdown compile

compile

Generate and save a readable `.enzyme` program. --preset is `arena`, `apple-notes`, or a path to a `.enzyme.in` template; remaining arguments are the preset source (Are.na channel URL, Apple Notes database path). --output writes a new file and never overwrites an existing one. --dry-run prints without saving. --json emits the program and the inventory it was selected from. --packet prints selection questions and evidence as JSON for an agent to answer; local only, saves nothing; conflicts with --preset, --output, --dry-run, and --json. --color is `auto` (default), `always`, or `never`. Markdown compile may call a model; source presets do not

Build or update the vault index and catalysts. -q / --quiet suppresses human output (for agents). --json-progress writes JSONL progress on stderr. --force / --regenerate discards the existing index and regenerates every entity's catalysts. --stage stops after a pipeline stage (for example `index`). --min-top-catalysts raises the ceiling for the largest entity. --lock-timeout SECS waits for another init/refresh of this workspace (default 1800; 0 fails immediately). --llm is documented under Catalyst generation

Incrementally refresh the index. Same -q, --min-top-catalysts, and --llm as init. --full / --force starts a complete catalyst rebase (still built detached). --lock-timeout defaults to 1800, or 5 with --quiet

Rank notes by keyword and meaning. Each result notes the connecting catalyst. -n / --limit defaults to 10. --candidates N lists up to N further hybrid matches (path, date, snippet). --routing is `hybrid` (default) or `catalysts`. --target DIR prepares and searches that directory instead of the vault and always routes through catalysts. --json always emits the JSON envelope, including when stdout is a terminal

Read-only search of the index: catalyst, direct, and exact-phrase hits. -n / --limit is the maximum per mode (default 10). --phrase is the literal phrase (defaults to the query). --json emits the `enzyme.search.v1` envelope

Show top entities and their catalyst questions. -n is how many entities to show (default 15; 0 means all). --query ranks entities and catalysts by relevance. --catalyst-budget is the maximum catalysts per entity in query mode (default 3). In a terminal this draws a tree; piped, it emits JSON

Show vault status and configuration. --json emits `enzyme.status.v1`

Report diagnostic facts. Read-only; never writes. --json emits structured JSON

Install or update this workspace's agent integration. TARGET is `claude`, `codex`, `hermes` (default), or `openclaw`. --source is `auto` (fetch GitHub copy, fall back to packaged; default), `remote`, or `local`. Claude and Codex skills go under the vault (`.claude/skills` or `.agents/skills`). Hermes and OpenClaw skills go under ~/.hermes / ~/.openclaw (or HERMES-HOME / OPENCLAW-HOME)

Sign in to memory.enzyme.garden. --no-open does not launch a browser. --json emits JSONL events for agent-mediated login. The first vault on a machine can init without login; refresh, publishing, account credits, and further vaults require this

Sign out of memory.enzyme.garden

Check for and install Enzyme binary updates. --check only reports, as JSON

Start an MCP server on stdio

Start the HTTP API for the explore UI. --port defaults to 8765

Download a GGUF into ~/.enzyme/models. Default name is `catalyst-135m-v3`. --url downloads a GGUF directly; --sha256 is the expected checksum for that URL

List installed and available models. --json emits `enzyme.models.v1`

Select an installed model and require local catalyst generation

Return to automatic selection (local when installed, otherwise hosted)

Disable local catalyst generation and use Enzyme hosted

Show the configured and effective catalyst-generation mode

Verify an installed model's file and registry checksum (defaults to the selected model)

Remove an installed model

Plan replacing workspace NAME's program with the text in FILE (`-` is stdin). Never writes the program; may finish an interrupted apply first. --json prints the plan; --out also saves it

Apply a saved plan. Refuses stale or altered plans; replays already-applied ones. --json prints the receipt. After a write, index with enzyme refresh --workspace NAME

Complete interrupted applies. --discard drops journals whose program was edited mid-apply, leaving the program as it is. --json prints the outcome

Parse and resolve a source file (no vault access, no model). --instructions or --normalized change the printout; default is the resolved plan as JSON

Export built-in profile definitions. --json emits `enzyme.profiles.v1`

Print the effective runtime config. --instructions prints the compiled agent policy instead

Preview what a program would do on an indexed vault (read-only, no model). Global -p / --workspace select the vault. --json, --limit N, --prompts DIR, --local (local-model prompts)

Migrate `config.toml` to `configs/*.enzyme`, keeping a timestamped backup. --home overrides the Enzyme home

Serve the syntax playground (Rust parser only). --bind defaults to `127.0.0.1`, --port to 4178

Ingest structured data into the Enzyme database. -q / --quiet emits JSON instead of human output

Embed structured entries without creating an Enzyme database. -i / --input is a JSON file (array, `{"entries": [...]}`, or `{"entry": ...}`); stdin if omitted

Benchmark retrieval. QUERIES is JSONL or plain text, one query per line. -n / --limit defaults to 10. --conditions is a comma list: `esedirect`, `esecatalysts`, `arcticdirect`, `arcticcatalysts` (default: all)

Dump catalyst prompts to JSONL. -o / --output defaults to `prompts.jsonl`. --min-top-catalysts as on init

init and refresh take --llm MODE:env

OpenAI-compatible endpoint from OPENAI-API-KEY (required), OPENAI-BASE-URL (default `https://api.openai.com/v1`), and OPENAI-MODEL (default `gpt-4o-mini`). Never Enzyme auth. --use-env-llm is an alias

The selected model in ~/.enzyme/models. Fails (exit 3) if none is installed

Enzyme hosted generation

Index, embed, and select only; no catalyst generation or network

Exit codes: 0 success (including warnings), 1 index or runtime failure, 2 usage (bad flags or workspace name), 3 configuration (unreadable program, undeclared workspace, missing --llm env / local prerequisites), 4 workspace busy (lock held past --lock-timeout), 5 init indexed but catalysts failed their quality gate.Concurrent init / refresh on one workspace wait on `workspace.lock`, bounded by --lock-timeout. A detached catalyst worker may hold `refresh.lock` separately.On FUSE mounts, the SQLite database is mirrored under `/tmp/enzyme/` because FUSE cannot support SQLite's `xDelete`.scan is a hidden command still used by the SQLite source skill (`enzyme scan --workspace` NAME); prefer compile for setup.This is Enzyme Labs' knowledge CLI, not the LLVM Enzyme automatic-differentiation plugin.

Enzyme is developed by Enzyme Labs (Joshua Pham). The public tree is a Rust workspace (`enzyme-cli`, `enzyme-core`, `enzyme-api`) under Business Source License 1.1, version 0.13.1 as of 2026-10-24. Releases and `install.sh` are published on GitHub.

rg(1), grep(1), jq(1), claude(1), codex(1), obsidian(1), sqlite3(1), llm(1)

Source code · Homepage · Documentation