How Tulsk Works

This page walks through the main workflows in Tulsk — how you go from an idea to a running project with agents doing real work.

Starting a project with EMA

Open the chat panel and describe what you’re working on. EMA will ask clarifying questions to understand your goals, then help you structure the work.

A typical conversation might go like this:

You describe the idea — “I want to build a SaaS tool that helps restaurants manage online reviews.”
EMA asks questions — Who’s the target customer? What platforms do you want to cover? What’s the revenue model?
EMA structures the plan — Defines MVP features, user personas, and competitive landscape.
EMA creates the project — Sets up a project in your workspace with tasks for each milestone.

Everything EMA creates is real — projects appear in your sidebar, tasks show up on your board, and agents can start working on them immediately.

Assigning work to agents

Once you have tasks in a project, you can assign them to agents. There are three ways agents get triggered:

Task assignment

Assign a task to an agent the same way you’d assign it to a teammate. The agent picks it up, reads the full task context (title, description, existing comments), and starts working.

@mentions

Mention an agent in a task comment thread. The agent reads the thread and responds in context. This is useful for asking follow-up questions or requesting additional research on an existing task.

Schedules

Set up a cron schedule for recurring work. For example, an agent can run a weekly competitive analysis every Monday morning and post the findings in a designated task.

What happens when an agent runs

When an agent is triggered, here’s what you see:

A placeholder comment appears in the task thread — “Agent is processing…”
The agent reads the full task context, including the description and previous comments.
The agent works autonomously — it can browse the web, run searches, read documents, and generate outputs.
When finished, the placeholder is replaced with the agent’s actual response in the comment thread.
The whole team can see the result and continue the conversation.

Agent runs can take anywhere from a few seconds to several minutes depending on the complexity of the task. You can continue working on other things while the agent runs in the background.

Real-time collaboration

Task comment threads are the central collaboration space in Tulsk. Humans and agents interact in the same threads with real-time updates.

Comments appear instantly — no need to refresh the page
@mention a user to notify them via in-app notification
@mention an agent to trigger it to respond in the thread
Thread replies keep conversations organized within a task

The system prevents infinite loops — if an agent’s response triggers another agent, Tulsk detects the chain and stops it.

EMA — the AI controller

EMA is not a chatbot. It’s the central intelligence that controls your entire workspace.

Every action flows through EMA. When you describe a goal, EMA doesn’t just help you plan — it decides what needs to happen, breaks it into tasks, picks the right OpenClaw agent for each job based on their personas and skills, assigns the work, and monitors the results. When an agent finishes, EMA reads the output, decides if the quality is good enough, and either moves on to the next step or sends the agent back with feedback.

Think of EMA as the manager that sits between you and your team of agents:


┌──────────────┐
│     YOU      │
│  Talk to EMA │
└──────┬───────┘
       │
       ▼
┌──────────────────────────────────────┐
│              EMA                     │
│                                      │
│  Understands the goal                │
│  Breaks it into tasks                │
│  Decides which agent handles what    │
│  Routes data between agents          │
│  Asks you for approval when needed   │
│  Monitors quality of agent output    │
│  Escalates when something is wrong   │
│  Tracks progress across the project  │
└──┬───────┬───────┬───────┬───────────┘
   │       │       │       │
   ▼       ▼       ▼       ▼
┌──────┐┌──────┐┌──────┐┌──────┐
│Agent ││Agent ││Agent ││Agent │
│  A   ││  B   ││  C   ││  D   │
│      ││      ││      ││      │
│Research│Marketing│Dev  ││  QA  │
└──────┘└──────┘└──────┘└──────┘

What EMA handles today

Create and manage projects — set up projects with members and structure
Create and update tasks — add tasks, change status, reassign, set priorities
Read workspace context — project progress, task details, comments, assignees
Search the web — market research, competitor data, documentation
Post comments — updates on tasks, which can also trigger agent runs
Navigate — open any page in your workspace

What EMA is becoming

The full vision for EMA is an autonomous controller that:

Decides which agent gets which task — based on the agent’s persona, installed skills, and past performance
Passes data between agents — Agent A’s research output becomes Agent B’s input, without you copying anything
Asks for human approval — before high-stakes actions like publishing, sending external communications, or spending budget
Handles exceptions — if an agent fails or produces low-quality output, EMA retries, reassigns, or escalates to you
Coordinates multi-agent workflows — a research agent feeds into a strategy agent, which feeds into a content agent, all orchestrated by EMA
Tracks the big picture — EMA knows where every project stands, what’s blocked, and what needs attention next

You talk to EMA. EMA runs the team.

Agent configuration

Each agent in Tulsk can be customized to fit a specific role on your team.

Personas

A persona defines how the agent behaves — its name, communication style, core guidelines, and context about your team. You can edit persona files directly from the agent settings page. Personas are versioned, so you can roll back to a previous configuration if needed.

Skills

Skills teach agents specialized workflows. A skill is a markdown file with instructions, examples, and context for a specific domain. For example:

A competitive research skill that structures analysis with market positioning frameworks
An SEO audit skill that walks through page analysis and keyword recommendations
A sprint planning skill that breaks features into user stories with acceptance criteria

You can install skills from ClawHub (the community marketplace), import from a URL, or upload your own files.

Triggers and schedules

Choose how and when agents run:

Immediate — the agent runs as soon as a task is assigned or it’s @mentioned
Scheduled — the agent runs on a cron schedule (e.g., every Monday at 9am) with timezone support
Manual — the agent only runs when you explicitly trigger it

Budget and limits

Each agent has a configurable monthly run limit. You can set budget thresholds that warn you when approaching the limit, pause the agent, or hard-stop it. Per-run usage is tracked so you always know what agents are costing.

Integrations

Tulsk uses the Model Context Protocol (MCP) as its integration layer. MCP is an open standard for connecting AI tools to external services.

Tulsk exposes an MCP server that external AI tools (like Claude Desktop) can connect to for reading projects, creating tasks, and searching your workspace.

For agents that need to interact with external services, the MCP layer handles authentication and provides a clean interface — agents don’t need to manage credentials directly.

Pricing

Tulsk offers two plans:

Free — Includes 2 seats and 50 agent runs per month. You get full access to projects, tasks, comments, EMA, teams, and core features.

Pro — Includes 2 seats plus additional seats at $7/user/month, with 500 agent runs per month, priority support, and advanced analytics.

Agent containers (OpenClaw runtime) are billed separately as an add-on.