Get started

This is the short path from “I just signed up” to “an AI agent is working on my issues”. Five steps, about ten minutes, on the Free plan, no card.

Each step keeps only what you need to reach the end. Where there’s more detail than fits here, there’s a link to the page that covers it in full.

1. Create your space

Sign up at app.hilbana.com/signup. Creating the account sets up your workspace with a General project and the six starting states (Backlog, Todo, In Progress, In Review, Done, Canceled).

You can rename them later, see Workflow states.

2. Write your first issue

You can work in the General project or create your own from the “+” button in the sidebar, next to “Projects”. Inside, go to the Issues tab and create one.

Besides the title and description, fill in the two fields that turn an issue into something an agent can execute:

• Definition of done: what has to be true for the work to count as finished.
• Verify command: what to run to check it, for example pnpm build or your tests.

Those two are the Definition of Ready. Without them an agent can touch the issue but has no way to know when it’s done. Also tick “agent ready” if you want it in the queue. Detail in Issues.

A small, real piece of your own work is fine for this tutorial. Note its identifier (DRAPPS-1, with your team’s prefix): you’ll use it in step 4.

3. Connect your agent

Hilbana speaks MCP (Model Context Protocol), so any MCP client can drive your board. Below are the three usual routes. Pick one, you don’t need all three.

Claude on the web (claude.ai)

The quickest connection, and the reference example for the rest: you sign in with your account in the browser (OAuth) and manage no keys at all.

1. In claude.ai, open Customize → Connectors.
2. Click + → Add custom connector.
3. Name it Hilbana, remote MCP server URL https://app.hilbana.com/mcp. On a self-hosted install, https://<your-host>/mcp.
4. Click Add, then Connect. Hilbana’s screen opens: sign in if needed, pick the workspace and the permissions, and click Authorize.

Grant the write permission if you want the agent to move states and comment, not just read.

The full steps, the permissions and how to revoke access are in Connect with Claude. That page is the reference; this is the summary.

Claude Code

In your terminal, inside the repository you work in. The plugin is the way to go here: on top of the MCP server it brings per-project memory, token accounting and the work-cycle commands:

/plugin marketplace add hilbana/claude-plugin
/plugin install hilbana@hilbana

Restart Claude Code. It asks for an API key (hil_…), created under Settings → API keys in the workspace you want to work in. Installation, setup and common problems in Claude Code plugin.

ChatGPT

ChatGPT does support third-party remote MCP servers, but only through developer mode: on the web (not the mobile apps), on Plus, Pro, Business, Enterprise and Education plans, and still as a beta feature.

1. In ChatGPT, turn on developer mode in settings. OpenAI has moved that toggle more than once: look under Settings → Apps and connectors → Advanced or Settings → Security. On Business or Enterprise accounts an admin has to enable it for the workspace first.
2. In the apps and connectors panel, click + and create a developer-mode app pointing at https://app.hilbana.com/mcp.
3. Choose OAuth authentication. Hilbana’s screen opens: sign in, pick workspace and permissions, authorize.
4. In the conversation, enable the app from the + menu before asking it anything.

Two warnings, so it doesn’t catch you out:

• The hil_… API key is no use here. ChatGPT won’t let you set custom headers on a custom connector: it takes OAuth or servers with no authentication. For Hilbana, OAuth.
• Custom apps are not verified by OpenAI, and the product says so itself. That’s expected for your own MCP server, but worth knowing.

Checked on 5 August 2026 against OpenAI’s documentation. It’s a beta feature and the menus get renamed often; if something doesn’t match, go by what’s on your screen.

Other agents

Any other MCP client (your own agent, a service, CI) connects with an API key and a config entry. That’s in API & MCP.

4. Your first prompt

Once connected, just ask in plain language. To start, something you can check at a glance on the board:

You're connected to my Hilbana over MCP. Work on issue [DRAPPS-N]:

1. Read it with get_issue, including the agent context and the definition of done.
2. Move it to In Progress with change_issue_state.
3. Post a comment with add_comment laying out your plan in three or four bullets,
   without executing it yet.
4. Tell me what the issue is missing for you to finish it without asking me anything
   else.

Replace [DRAPPS-N] with the identifier from step 2. By default your client will ask for approval before using each tool.

On the board you should see the issue in In Progress and the comment signed by your user. That last bullet is the one that pays off early: it tells you what your issue is missing to be executable.

Once this works, the Prompts catalogue has the full flows: kick off a project, turn an idea into an epic, pull from the queue or review what sits in In Review.

5. Where to go next

• Connect with Claude: the claude.ai connector in full, permissions and revocation.
• Claude Code plugin: memory, cycle commands and token accounting.
• API & MCP: keys, endpoint and every MCP tool.
• Prompts: the catalogue of ready-to-copy flows.
• Agents: the queue, claim/release and the live panel.
• Issues: how to write issues an agent can execute.