A cross-platform desktop command center for running and supervising AI agent teams — built and tested on macOS first.
-black){kind=icon}
Codenest is a local-first desktop app for running and supervising teams of AI coding agents. It hosts embedded terminals that run real Claude Code CLI sessions, tracks their work on a Work board, and runs agents on cron Schedules — all from a single window, with every bit of state kept in a local SQLite database on your machine.
It is built on a cross-platform stack (Tauri 2, a Python sidecar, and a React frontend) intended for macOS, Windows, and Linux. Today it is built and tested on macOS (Apple Silicon) — the only supported platform for now; other platforms are on the roadmap.
Grab the latest macOS (Apple Silicon) .dmg from the
Releases page. Builds for
other platforms will be added there as they are tested.
-
Open the
.dmgand drag Codenest.app into/Applications. -
Releases are unsigned, so clear the quarantine attribute:
xattr -cr /Applications/Codenest.app
-
On first launch, if macOS blocks the app: macOS 15+ open System Settings → Privacy & Security and click Open Anyway; on older macOS, right-click the app and choose Open.
See docs/install.md for the full walkthrough, where data lives, and how to reset the app.
Codenest's rail is organised into four groups. Needs You, Work, Schedules, Budgets and Hooks can each be switched off under Settings → Features.
Now
- Deck — the home screen: what needs you, what is running, what it is costing today, and what changed since you last looked.
- Needs You — the intervention queue: stalled sessions, failed scheduled runs, budget thresholds, blocked tasks and unread notifications in one list. Every row offers the one action that gets you to the problem — focus the live pane, open the session, start one seeded from the ticket, or jump to the surface that owns it.
- Work — tasks grouped by status, with a detail page carrying subtasks, comments, blockers, activity, agent runs and what the task has cost.
- Sessions — two halves of one surface: panes, where a session runs (native agent panes and PTY-backed xterm.js terminals), and runs, where every run is listed, filtered, focused, stopped and inspected.
Record
- Projects — the repositories your agents work in, each with a context panel showing what ran there and what it cost.
- Agents — your team of agent definitions, and what each has been invoked for.
Automate
- Schedules — cron- and event-triggered agent runs, with their run history and transcripts.
System
- Budgets — usage and cost budgets by workspace or project.
- Hooks — what is on each Claude Code hook event, and which of the eight possible sources put it there.
- Settings — app, provider, profile and feature configuration.
Needs You — the intervention queue. Each row names what is wrong, where it is, how long it has waited, and offers the single action that takes you there.
Sessions → runs — every agent run in one list, with its provider, project, model, current tool, elapsed time and cost, plus a live activity feed. Focus jumps straight to the pane the run is in; it is the same page.
Work — tasks grouped by status, filterable by project and assignee.
Launch a session — pick a recipe or a saved preset, then edit an ordered list of agent and shell panes, each with its own provider, model and permission mode. One shared prompt feeds every agent pane that opts in; choose a project, profile and target, then launch or save the composition as a preset.
More screenshots — Projects, Agents, Schedules, Budgets, Settings, first run
Projects — the repositories your agents work in. Imported projects stay read-only sources; nothing is ever written into their folders.
Agents — your team of agent definitions, grouped by scope. Codenest ships with three org agents (Atlas, Orion, Vega) shared across the workspace.
Schedules — cron- and event-triggered runs, with their history and transcripts.
Budgets — per-workspace or per-project spend caps by period, with an optional hard stop that blocks new sessions at 100%.
Settings — providers, profiles, workflow vocabulary, terminal, telemetry and the feature toggles.
First-run setup — a seven-step wizard that creates the app-managed workspace, imports your projects and connects your AI provider.
A Tauri 2 (Rust) shell is the process supervisor. On launch it spawns a FastAPI
(Python) sidecar that owns all application state in SQLite and serves it on
127.0.0.1:8002, then waits for the sidecar's health check before showing the
UI. The React/TypeScript frontend talks to the sidecar over HTTP/SSE and uses
Tauri invoke for shell-level work — PTY-backed xterm.js terminals, windows,
the filesystem watcher and git. Agent process spawning lives in the Rust
shell; the sidecar queues due scheduled runs and the shell dispatches them.
Everything runs on your machine: the sidecar is loopback-only and the app sends
no telemetry.
Prerequisites:
- macOS (Apple Silicon) — the only platform tested so far
- Xcode Command Line Tools
- Rust (stable, 1.88+)
- Node 20+ with pnpm
- Python 3.12
python3 -m venv .venv && .venv/bin/pip install -r requirements.txt
pnpm install
pnpm tauri:dev # sidecar archive auto-builds via the Tauri beforeDevCommand hook
pnpm tauri:build # production .app + DMGThe Python sidecar is packaged into an archive by scripts/build-sidecar.sh.
You do not need to run it by hand: pnpm tauri:dev builds it on first run if
missing, and pnpm tauri:build always rebuilds it. To prebuild it manually, run
bash scripts/build-sidecar.sh.
Contributions are welcome — see CONTRIBUTING.md for setup, the
make check-all gate, and project conventions.
Codenest is licensed under the Functional Source License, Version 1.1, ALv2
Future License (FSL-1.1-ALv2) — a source-available license. You
may use, copy, modify, and redistribute the source for any purpose except a
competing commercial use. Two years after each version is released, that
version automatically converts to the Apache License 2.0.
See NOTICE for attribution and THIRD-PARTY-NOTICES.md for bundled third-party dependencies.