Skip to Content
Docs are being rebuilt — start at Introduction → How it works.
IntroductionHow it works

How it works

TeamMate has six building blocks. Get the names right and the rest of the docs make sense.

The picture

┌──────────────────── Workspace ─────────────────────┐
│                                                    │
│  ┌─ Agents ──────────────────────────────┐         │
│  │   instructions · model · personality  │         │
│  │   ┌────────┐  ┌────────┐  ┌────────┐  │         │
│  │   │  Atlas │  │Mutarjim│  │   …    │  │         │
│  │   └───┬────┘  └───┬────┘  └───┬────┘  │         │
│  └───────┼───────────┼───────────┼───────┘         │
│          │           │           │                 │
│  ┌───────▼───────────▼───────────▼──────┐          │
│  │  Knowledge · Tools · Skills          │          │
│  │  (what each agent can read & do)     │          │
│  └──────────────────┬───────────────────┘          │
│                     │                              │
│  ┌──────────────────▼───────────────────┐          │
│  │  Channels  ·  Interfaces             │          │
│  │  (how users reach the agent)         │          │
│  └──────────────────┬───────────────────┘          │
│                     │                              │
│  ┌──────────────────▼───────────────────┐          │
│  │  Triggers · Tasks · Workflows        │          │
│  │  (when the agent runs by itself)     │          │
│  └──────────────────────────────────────┘          │
│                                                    │
└────────────────────────────────────────────────────┘

Read it top-down: a workspace holds agents; each agent has knowledge / tools / skills; users reach agents through channels and interfaces; triggers run agents on their own.

The six blocks

1. Workspace

Your team’s private space inside TeamMate. Everything in it — agents, conversations, files, members, credits — belongs to one workspace. Each company is one workspace. Owners and admins manage it. → Workspace

2. Agent

A named AI employee with a job, instructions, and a personality. Atlas (technology Q&A), Hashim (sales coordination), Mutarjim (Arabic ↔ English translation). Every agent is built from a clear instruction document, optional knowledge bases, optional tools, and a chosen model. → Agent

3. Knowledge

Documents the agent can search. Policy PDFs, product specs, FAQs, runbooks. When the agent gets a question, it searches its attached knowledge bases first, then composes a reply with the relevant sources cited inline. → Knowledge base

4. Tools (and Capabilities)

What the agent can do, in addition to talking. Tools include web search, code execution, file management, sending email, talking to Microsoft Teams or Slack, querying a database, controlling a browser. A capability is the higher-level switch (“this agent can use web search”) — TeamMate decides which specific tool fits the model and the task. → Tool & capability

5. Channels and Interfaces

How users actually reach the agent. The default is chat in the web app. Beyond that, agents can be put on channels — Slack, Teams, WhatsApp, email, voice (Twilio, Vapi), or an embeddable web widget. Interfaces are pre-built forms layered on top of the chat (e.g. a “report an incident” form that sends a structured message to an agent). → Channel · Interface

6. Triggers, Tasks, and Workflows

How agents run without a user typing. A trigger is the event that starts a run — a schedule, an incoming webhook, an email landing in an inbox, a file being uploaded. A task is one agent running on a schedule. A workflow is a multi-step recipe that connects triggers, agents, and integrations — useful when one agent isn’t enough. → Task vs workflow vs trigger

How a single interaction flows

When a user sends “summarise yesterday’s tickets” to the HR Assistant:

  1. The message lands in the agent’s chat surface (web, Slack, Teams — wherever it was sent).
  2. The agent reads its instructions and your message.
  3. If the question references documents the agent has access to, it searches its knowledge bases and pulls relevant chunks.
  4. If it needs to do something (run code, call an API, look at a database), it uses one of its tools.
  5. The chosen model drafts a reply, streamed back to the user word by word.
  6. The whole turn — message, tool calls, model reply — is recorded in the workspace’s audit log.

The “live” stream you see in chat is the agent thinking out loud and then producing the final reply. Approvals can intercept the agent before any destructive action runs.

Where governance fits

Every block has a governance hook:

  • Workspace → role-based access (owner / admin / member), workspace-level audit log.
  • Agent → versioned instructions, optional human approval before sensitive tools fire.
  • Knowledge → per-document access control; embeddings stay on TeamMate infrastructure unless you connect an external vector store.
  • Tools → enabled per agent. Sensitive tools can require approval.
  • Channels → channel-level allow-lists (e.g. only this Slack workspace can call this agent).
  • Triggers / Tasks → owner attribution, pause / resume, last-N-runs audit history.

Read Governance for the full picture.

Where the model fits

TeamMate is not wedded to one model. Each agent picks its own model at the AI Model tab in the agent builder. Choices today include OpenAI (GPT family), Anthropic (Claude family), DeepSeek, OpenRouter, and any OpenAI-compatible custom endpoint — useful for on-prem deployments (vLLM, LiteLLM, or a model hosted by your IT team).

The runtime adapts: if an agent’s chosen model can’t read a PDF inline, TeamMate auto-routes the PDF through a text-extraction tool. If a model doesn’t natively speak Arabic well, you can pick one that does. This is what we call capability-aware routing.

Where the workspace’s data lives

  • Conversations, messages, agent configs, knowledge metadata → TeamMate’s primary database (PostgreSQL).
  • Knowledge embeddings → either TeamMate-managed (default) or your own vector store (S3-backed snapshots, Pinecone, Vespa).
  • Workspace files → S3 (or DigitalOcean Spaces, depending on deployment).
  • Workspace secrets (API keys for tools, connected app tokens) → encrypted at rest using AES-256-GCM with a workspace-scoped key.

Where to next

What is TeamMate?Quick start (5 min)