@jgalego/teamapi-chat

@jgalego/teamapi-chat

Chat as a team or a specific team member from a Team API as Code org — backed by a live tool-use loop over the same ~12 org-graph operations @jgalego/teamapi-mcp-server exposes, so the persona can accurately answer questions about any team, not just its own.

Normally used via teamapi chat --team <id> [--member <id>] [--ask <question>].

Providers

Two adapters, which is what the interoperability landscape actually has:

provider	endpoint	key
anthropic	the Anthropic Messages API	ANTHROPIC_API_KEY, required
openai	any OpenAI Chat Completions server	OPENAI_API_KEY, optional

The openai adapter is fetch against a configurable base URL rather than a vendor SDK, because that wire format is the de facto interoperability layer: Azure OpenAI, Ollama, vLLM, llama.cpp, Together, Groq, Fireworks, OpenRouter and most self-hosted gateways all speak it. The key is optional so a model running locally — which has none — is not the one case that fails.

Install

npm install @jgalego/teamapi-chat

Usage

import { buildChatPersona, buildChatTools, createChatSession } from "@jgalego/teamapi-chat";

const persona = buildChatPersona(graph, { teamId: "stream-checkout", memberId: "diego-alves" });

const session = createChatSession({
  provider: "openai",
  baseUrl: "http://localhost:11434/v1", // or omit for api.openai.com
  model: "llama3.1",
  system: persona.systemPrompt,
  tools: buildChatTools(graph),
});

const answer = await session.ask("is payments overloaded right now?");
console.log(answer.text);
if (answer.stoppedBecause) console.warn(`incomplete: ${answer.stoppedBecause}`);

ask never pretends a turn finished when it didn't: stoppedBecause is tool-limit, refusal or truncated when the model stopped for a reason other than completing its answer.

Adding a provider

A tool is a name, a description, a zod schema and a function — ChatTool, in tool.ts, with no vendor content in it. An adapter converts that list to the provider's shape and drives the call-observe-call loop, which is about a hundred lines. runToolByName handles the parts every adapter needs identically: validating arguments against the schema, and turning an unknown tool or a thrown error into a message the model can read and recover from rather than an exception that ends the turn.

Full docs and a sample transcript: https://github.com/JGalego/TeamAPI

The TeamAPI toolchain

One org graph, seven doors into it — install only the ones you need:

Package	What it does
@jgalego/teamapi	The CLI — validate, diagram, check, import, reconcile, serve and chat with your org
@jgalego/teamapi-core	The engine: $ref resolution, the org graph, scoring, checks, diagrams, generators
@jgalego/teamapi-schema	Zod schemas and TypeScript types for the extended spec
@jgalego/teamapi-rest-api	REST API, live dashboard, Swagger UI, Prometheus metrics
@jgalego/teamapi-mcp-server	The org graph as MCP tools for LLM assistants
@jgalego/teamapi-chat (this package)	Chat as a team or member — Anthropic or any OpenAI-compatible endpoint
@jgalego/teamapi-backstage	Live Backstage catalog entity provider

Docs, examples and the extended spec: teamapi.dev · github.com/JGalego/TeamAPI

License

MIT