TinyAGI

Why it's ranked here

The mechanism here is real: a working SQLite queue, per-provider adapters for Claude and Codex, a documented plugin and hook system, and a chat-room protocol that actual code implements, not just describes. That earns a look. But the project's own badge calls it experimental, and the way it invokes the underlying coding-agent CLIs turns off their permission prompts and sandboxing entirely, trading safety checks for hands-off automation. That is fine on a machine you already trust and risky anywhere else. Worth studying the architecture; worth hesitating before letting it run unattended on anything that matters.

What's good

The queue is built like production infrastructure, not a demo: retries with a capped count before a message is marked dead, automatic recovery for anything stuck mid-process after a crash, and separate append-only tables for chat history and per-agent messages. Provider support is a small adapter registry, so adding a new backend means writing one adapter, not touching the core. Plugin hooks that reshape incoming or outgoing text are isolated: the docs state a failing plugin cannot crash the queue processor. The agent handoff protocol, plain mentions in a reply that become new queued messages, is simple enough to reason about without reading the source.

Tradeoffs

The agent invocation deliberately turns off its own safety rails: the underlying coding-agent CLIs are called with flags that skip permission prompts and bypass the sandbox, so an agent that goes off script has no guard between it and the filesystem. Two history tables, chat messages and per-agent messages, are documented as append-only and growing indefinitely, with no pruning described for either, unlike the message queue's own cleanup of completed and acknowledged entries. The project labels itself experimental on its own badge, and the optional personality-customization file ships as a bracketed fill-in-the-blank template rather than a working default.

How to use it well

This fits someone who wants a fleet of coding agents reachable from chat apps, each in its own workspace with its own memory, handing tasks to each other without a human relaying messages between them. Run it on a machine you already trust with full CLI access, not a shared or public one, since the underlying agents' permission prompts are switched off by design. Plan to prune or archive the growing chat and per-agent history yourself, since the built-in cleanup only touches finished queue entries, not those logs. Skip it if you need agents constrained by a sandbox or an approval step before they touch files.

Observed

License

MIT (LICENSE file)

Language & packaging

TypeScript, npm workspaces monorepo with a single tinyagi CLI binary

Install surface

Shell install script, npm install from source, or Docker Compose

Interfaces

CLI, REST + SSE API (default port 3777, configurable), browser dashboard, and Discord/Telegram/WhatsApp channel bots

Platform support

macOS, Linux, and Windows via WSL2; requires Node.js v18+

Agent execution

Shells out to the Claude and Codex CLIs with permission-prompt and sandbox flags disabled

Persistence

SQLite-backed queue; chat and per-agent history tables documented as append-only with no pruning