Minnow

Loading…

Minnow
~…

General / Startup & setup

Startup & setup

Terminal

Data & privacy

Updates

Notifications

Audio

Devices

About

Plugins

GitHub

Issues

Appearance

Providers

Routing

Sampler

Thinking

Usage & cost

Voice

Agents

System prompt

Shared Full, Lite, and Custom instructions applied before mode and agent prompts.

Computing estimate…

Chat context

Rules

Agent packs

Bundled agent definitions you can enable for a workspace or session.

Autopilot

Timeouts & recovery

Web search

Servers

Tool permissions

Skills

Skills Library

Browser

MCP servers

Connect other apps

Language servers

Editor

Webhooks

Send signed notifications to another service when chats, sessions, or scheduled jobs change.

Health & diagnostics

Choose a workspace

New projects save to your default folder

Recent workspaces

    What should we look into?

    Minnow reads the web, or your codebase, and writes back a brief with its sources.

    Try

    Chat

    Models

    Engine

    The runtimes that serve local models. Choose upstream llama.cpp or a fork, and manage MLX on Apple Silicon.

    Storage

    Where Minnow looks for weights, and the token it uses to fetch gated ones.

    Knowledge graph

    Explore wiki pages, tags, and wikilinks on an interactive canvas.

    Markdown source

    Rendered page

    Untitled page

    —

    Start typing in the source pane to see a live preview.

    Read-only history from log.md.

    Edit schema.md taxonomy for tags and frontmatter fields.

    schema.md

    Live reference

    Edit schema.md to define wiki taxonomy fields and tags.

    Pending: —

    Prompt injection defaults are managed in Settings → Agents → Injection.

    — entries Checking server…

    Select an entry to read the full text, or add a new memory.

    Paste raw sources; the utility model writes wiki pages under brain/pages/.

    Source

    Read-only wiki scan, then an AI markdown plan you review before any writes.

    Checking index…

    100%

    Checking wiki status…

    Auto-learning cadence

    Runs after every 4 completed user+assistant turns.

    Off by default. Board runs already synthesize once per task at build pass and after merge; turning this on adds the throttled per-turn pass for planner, builder, tester and fixer chats.

    Facts at or above this confidence are written to the wiki directly. Below it, they queue as proposals for review.

    Skill proposals

    Propose a skill when separate sessions solve the same kind of problem.

    Set sessions to 1 to propose from a single session.

    Semantic embeddings

    Checking embeddings…

    Vector blend 50% vector

    Code index

    Checking code index…

    File patterns

    Token budgets

    Injection budget applies to per-send composer code-map injection. Repo map budget is for the tool and Brain Code map panel.

    Composer & reindex

    Advanced

    Semantic code search is not wired yet; this toggle is saved for a future release.

    Indexing scaffold lives under .minnow/ and is not written to the repo root. Disable if you manage project config yourself.

    Danger zone

    Destructive actions cannot be undone. Backup first when offered.

    Wiki

    Remove all wiki pages or pending memory and skill proposals.

    Ingest sources

    Delete raw ingest files. Wiki pages are kept.

    Code index

    Reset indexed symbols for this workspace or every workspace.

    Applies to every workspace on this machine.

    History

    Win rates

    Win rates appear after your first vote.

    Scheduler

    —

    Experts' Lab

    Compose, tune, and test expert agents

    Describe the specialist you want. Minnow will draft a custom expert you can reuse in Experts' Lab.

    Update your custom expert's details and prompt.

    No messages yet

    Pick a model above, then type below. LM Studio must be running at the server URL in Settings.