Value proposition

What you get when you adopt agents-in-a-box.

The promise

Run agents. Worktree per session. No cross-contamination.

That’s the whole pitch. The rest is what makes it real:

What you get

Multi-provider session management in one tool

Spawn Claude Code, Codex, Gemini, Copilot, Kiro, or a raw shell session from the same interface. Model picker per session (Sonnet · Opus · Haiku). Keyboard-driven. No web dashboard.

Git worktree isolation by default

Every session gets its own branch and worktree directory. No stash dance. No cross-contamination between parallel sessions. Auto-cleanup when sessions are killed cleanly.

tmux-backed persistence

Sessions survive terminal disconnects, SSH drops, laptop sleep. Reattach any time with ainb attach <name> or via the TUI.

Built-in usage analytics

Token and session attribution per day, week, project, model, and provider. Burndown dashboard with optimisation hints. No external service needed — reads local JSONL logs.

Scriptable end-to-end

Every TUI operation is also a CLI subcommand. --format json on every command. Pipe to jq, drive from CI, automate workflows.

A portable toolkit that follows you across tools

86 skills (plan, implement, validate, reflect, swarm-create, …) and 37 specialised agents (backend-developer, code-reviewer, security-agent, …). Write them once; deploy to Claude Code, Codex, Copilot, Gemini, Amazon Q, Cursor, Cline, Roo, Hermes, nanoclaw, Clawdhub. One source, 11 targets.

A plugin system you can actually use

Add a TUI screen, a CLI subcommand, a statusline segment without forking the host. Native binaries over JSON-RPC. Capability-gated — your plugin can’t reach the network unless you say it can. Two reference implementations ship in-tree; copy-paste a starting point.

A knowledge system that learns across sessions

/reflect captures non-obvious learnings as you work. /recall retrieves them in future sessions. GraphRAG underneath for cross-project queries; QMD vector search for fast hits. 170+ learnings indexed across the maintainer’s own dev work, growing.

What it costs you

One install command. brew install ainb (macOS, Linux). One-liner install.sh for everything else.
No subscription. MIT-licensed. No accounts, no metering.
No vendor lock-in. Toolkit deploys to 11 AI tools; the TUI orchestrates four providers. You can swap out any layer.
No cloud dependency. Everything runs locally. The only network traffic is what your AI provider (Claude, Codex, etc.) makes on your behalf.

What it’s worth

If you’re running ≥2 parallel AI coding sessions a week and tracking your own token spend, agents-in-a-box pays for itself in the first week — purely on saved worktree-management toil. The burndown dashboard alone replaces a dashboard you’d otherwise build yourself (or live without).

If you’re authoring AI skills or agents and supporting them across multiple tools (Claude Code + Codex + Copilot), the toolkit’s deploy-everywhere bootstrap replaces nine custom packaging scripts.

If you’re building a TUI extension or a custom analytics dashboard for your team’s AI usage, the v2 plugin system lets you ship that as a 500-line binary instead of a fork.