What it is
A small TypeScript library that sends each AI conversation in a product to a Slack thread. The team sees each message, tool call, and response.

Why
When an AI feature is new, the fastest way to learn is to read real conversations. Most tracing tools put them in a dashboard that nobody opens.
Slack is where the team already is. A conversation that appears there gets read.
What it does
Each conversation becomes one Slack thread. The first message shows the user. Each event is a reply:
- the message of the user
- the response of the assistant
- each tool call, with its arguments
- each tool result
- the reasoning of the assistant
- errors
How it works
Start a trace, log the events, and end the trace.
const trace = await bc.trace({ userId: "user_123" });
await trace.userInput("What's the weather?");
await trace.toolCall("get_weather", "call_1", { location: "NYC" });
await trace.toolResult("get_weather", "call_1", { temp: 72 });
await trace.assistantResponse("It's 72°F in NYC!");
await trace.end();
With the Vercel AI SDK, there is less to write. Wrap streamText or generateText, and the trace fills in with no other code.
A trace goes to one or more sinks. There are two: Slack and PostgreSQL.
Set it up
npm install breadcrumb-chat
npx breadcrumb slack
The second command creates the Slack app and saves the config.