OpenTag

Why it's ranked here

It ranks here because it is a complete, licensed, deployable system, not a demo: container images ship for both halves, deployment paths exist for two different hosts, and the code itself documents real production incidents it was patched to fix, from a middleware ordering bug that broke browser auth to a race in resuming paused approvals. That is unusually candid engineering for a repository meant to be forked, and it signals a team debugging this in a live workspace rather than only in a demo.

What's good

The approval flow for risky actions is genuinely durable: a decision made by clicking a button is stored so it survives a restart, and clicking it twenty minutes later still resolves the exact write it was meant for, not a random one, because each card is bound to the specific paused step that created it. The code also distinguishes carefully between a personal connected account and a shared workspace one for every action, and separates a lookup failure from an app the user genuinely has not connected yet, so error messages sent to a person are specific rather than generic.

Tradeoffs

Getting it running needs several outside accounts and services at once: a CopilotKit account, an OpenAI key, a Slack workspace with admin access, and by default a hosted service that handles delivery and state rather than a fully self-managed one. The stack itself is two languages glued together, a Node process and a separate Python process sharing one configuration file, which is more moving parts than a single-runtime bot. Only Slack and Microsoft Teams are supported chat platforms today; other messaging apps are not yet an option.

How to use it well

This suits a team that already has, or is willing to build, its own agent logic and wants a proven Slack or Teams wrapper around it: replace the example agent while keeping the channel plumbing, deployment scripts, and approval UI as-is. It is not the right pick for a team that wants a bot to install and configure through a UI, since running it means editing source, managing two runtimes, and standing up accounts with several outside services first. Best used as a template to fork and thin out, not as a dependency to import.

Observed

License

MIT

Polyglot codebase

a TypeScript Node process and a separate Python process (FastAPI plus LangGraph) that share a single root .env file

Packaging

Packaged as two separate container images, one for the agent process and one for the runtime process

Packaging

Two documented deployment targets: a two-service Railway configuration and a single AWS ECS Fargate task

Interfaces

Slack and Microsoft Teams chat, plus a small HTTP API (a health-check route and an account-connect route) alongside the chat surface

Structural

Writes to connected third-party tools are gated behind a human-in-the-loop approval interrupt backed by a durable, restart-surviving decision store rather than executing directly

Security

An optional shared-secret HTTP middleware protects most agent routes and is inactive when no secret is configured