aizen

An offline-first, local AI toolkit in Go. A single static binary that runs a retrieval-augmented knowledge index, project-scoped Q&A, and a set of deterministic utilities against local open-source models (via ollama) or any OpenAI-compatible endpoint. Proprietary APIs are optional accelerators — nothing critical depends on them, and most commands work fully offline.

aizen writes only under ~/aizen/. It does not modify your project repos or system files.

What it does

• ask — ask a local model directly, or with --project retrieve grounded answers from that project's index (hybrid BM25 + cosine RAG, with source citations).
• scan — zero-tolerance secret/PII scanner for a directory; --git-only restricts it to git-tracked files. Useful as a pre-push hook.
• engine — manage the local model layer: install ollama, list/pull models, set the default and the --deep batch model.
• pack / corpus — build knowledge packs from a repo + git log + memory, ingest research PDFs, and distill conversation transcripts into redacted, scope-tagged cards.
• model — model-assisted writing helpers: humanize text, generate commit messages and changelogs, check memory pointers.
• bb — drive an external bug-bounty toolkit safely (plan/run/explain) without modifying it; active scans are gated behind --authorize and a sandbox.
• run / loop / keyword macros — orchestrate subcommands concurrently, repeat them on an interval or until they pass, and bind short keywords to command chains.
• chat — an interactive REPL with editable, replayable history.
• skill / plugin — declarative extension points loaded from ~/aizen/.
• doctor — read-only health check of the local setup.

Global flags: --no-model (deterministic core only, never calls a model), --json (machine-readable output), --deep (route model calls to the big model).

Build

go build ./cmd/aizen

This produces an aizen binary. Go 1.25+ is required (see go.mod).

Usage

# One-time: install ollama and pull the open-source model ladder
aizen engine setup

# Health check
aizen doctor

# Scan a directory for leaked secrets/PII
aizen scan ./my-repo
aizen scan --git-only ./my-repo

# Ask the local model
aizen ask "explain this error message: ..."

# Register projects in ~/aizen/projects.conf, then query one:
#   alpha = /home/you/code/alpha
aizen ask --project alpha "where is auth handled?"

See docs/GUIDE.md for the full command reference, project and scope configuration, and orchestration primitives.

Configuration

• ~/aizen/projects.conf — register project names and their source repos (name = /path per line).
• ~/aizen/scopes.conf — define scope keywords for corpus distill (scope: kw1, kw2 per line).
• ~/.config/aizen/keywords.toml — define your own keyword macros.
• Environment: LLM_MODEL / LLM_BASE_URL select the model backend; AIZEN_HOME overrides the data directory.

License

MIT — see LICENSE.