Skip to content

NanoTeams 1.7.0 — Agent Skills & Instructions

Choose a tag to compare

View all tags
@jmstajim jmstajim released this 12 Jul 14:58
· 21 commits to main since this release
1.7.0
4cb0f60

✨ Highlights

NanoTeams now plugs into the agent tooling you already have. Work folders auto-pick up agent instruction files (CLAUDE.md, AGENTS.md, and friends) and feed them into every role's system prompt. A new "/" skills picker in the composer surfaces skills and slash-commands from Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, Windsurf, and OpenCode. And when a draft needs work, ✦ Improve Prompt streams a rewrite from your local model into the field.

🆕 Added

  • Agent instruction files — auto-pickup. NanoTeams discovers CLAUDE.md, AGENTS.md, GEMINI.md, .cursorrules, .github/copilot-instructions.md, and .windsurfrules anywhere in the work folder, injects the main file in full into every role's system prompt, and lists the rest as paths that roles can read on demand. A new grid in Settings → Work Folder lets you attach extra files, exclude a file from injection (it stays listed as a path), or promote any listed text file into the prompt. An "All files" card adds QuickLook over the whole set, and the Work Folder Context moves to a full-size editor sheet.
  • "/" skills picker. A "/" button next to "+" on every composer opens the skills and slash-commands found in the work folder and your home directory: Claude Code (including enabled plugins' skills and commands), Codex, Cursor, Gemini CLI, GitHub Copilot, Windsurf, and OpenCode, each labeled with its origin. Picking one stages a removable chip; on submit the skill's full content lands in the LLM prompt as a ## Skill: section. Hover any row for the full name, description, and path.
  • ✦ Improve Prompt. A sparkles button on every composer and on the Autovisor Goal fields streams an LLM rewrite of your draft live into the field, Apple Writing Tools style. The field locks while streaming, the button becomes a stop control, and a Revert chip restores your original. If the rewrite fails, you keep your original text and the button offers a retry.
  • Autovisor Goal — a real composer. All three Goal editors (setup pane, Settings, Watchtower card) now use the shared composer: file attachments stored as durable files that survive manager delete/recreate, clips, skills, dictation, and an "Embed files in prompt" toggle.
  • Autovisor — team-generation toggle. New per-folder setting "Let Autovisor generate new teams" in Settings → Autovisor. Off removes generation from the Autovisor's tool catalog and blocks it at runtime, so it works with existing teams only.
  • search & list_files — list mode. search with an empty query plus a file_glob or paths now returns the matching file roster without reading content, and list_files gains a name_glob filter, so "list every *.gd recursively" takes one call. An invalid glob returns a corrective error rather than a silent zero.

🛠 Improved

  • Work-folder context that fits your model. Context generation now probes the loaded model's context window and sizes its prompt to fit: the planner collapses large directories into aggregate lines, trims excerpts, and retries once with a smaller budget on overflow. This fixes generation on large, asset-heavy folders, and real server errors now surface instead of a generic banner.
  • Settings sidebar reorganized. The 7-tab "Advanced" grab-bag is gone: 17 tabs now sit in five task-shaped sections (Workspace, Models, Agent Tools, Application, Support), with Updates pinned at the top. Your remembered tab selection and deep links survive.
  • Vision on by default. New installations get image analysis without a settings trip.
  • Team generation sturdier on small local models. The create_team parser now repairs a dropped } / ] in the model's JSON (qwen3.5-35b: 7/11 → 11/11 tasks), artifact names stay conceptual deliverables instead of file names, and production tasks no longer collapse into Chat mode.
  • Tidier Bash approval advice. The AI second-opinion card no longer repeats the command above the verdict.

🐛 Fixed

  • Autovisor wakes on every task-status change. A task that failed before its engine existed (a failed team generation, say) used to wait for the next poll, up to a minute, before the Autovisor reacted. Now any status change wakes the manager.

💻 System Requirements

  • macOS 15.0 or later
  • LM Studio 0.4.0 or later
  • Apple Silicon recommended for best local-LLM performance
  • macOS 26+ required for private voice dictation

📦 Install

  1. Download NanoTeams.app.zip and extract it.
  2. Drag NanoTeams.app into /Applications.
  3. If macOS blocks the app on first launch: System Settings → Privacy & Security → Open Anyway.
  4. Launch LM Studio, load a recommended model, then launch NanoTeams.

🏗 Build from Source

git clone https://github.com/jmstajim/NanoTeams.git
cd NanoTeams
xcodebuild -project NanoTeams.xcodeproj -scheme NanoTeams -configuration Release build

🔗 Links

Topics: macOS · AI coding assistant · local LLM · LM Studio · multi-agent AI · AI agents · agent skills · computer use · desktop automation · screen control · accessibility · sandbox · security · automation · open source · private AI · on-device · offline AI · Apple Silicon · Swift · SwiftUI · agentic chat · code generation

Assets 3
Loading