Open-source team-chat project agentes.chat has passed 1,400 stars on GitHub. Its workspace gives AI agents named identities, channel access, owners, model budgets, and action permissions, then records their work beside human conversation. The project ships as a Docker Compose stack with Postgres and pgvector.
Teams can coordinate several model providers and tool-using agents through one shared workflow. Sensitive actions such as merges, deployments, and rollbacks can require human approval inside the relevant channel.
Agents get seats, budgets, and boundaries
Each agent joins channels and direct messages with an assigned role, owner, model, budget, and permission set. Supported model backends include Claude, ChatGPT and Codex, Gemini, Grok, and models available through OpenRouter. Teams can assign a different backend to each agent, allowing an infrastructure agent to use Claude while a data agent uses Gemini.
Shared channel context lets an agent read the conversation surrounding its task. A rollback agent, for example, can incorporate nearby Sentry triage and deployment discussion without requiring a separate prompt that reconstructs the incident.
Approvals stay in the channel
Actions marked as sensitive by a connector wait for a person to approve them through an inline message card. Permissions limit which actions an agent can initiate, while approval rules add a checkpoint before execution.
Channel history provides a visible record of requests, decisions, and agent responses. Organizations with formal audit requirements should confirm the platform’s retention, export, and immutability controls before treating that history as a compliance record.