What an agent can do
How a coding agent connects to Entalpa and signs in as you, what it can read, change and delete in your projects, and what stays your call.
How an agent connects
A coding agent, such as Claude Code, Codex or Cursor, connects to Entalpa's MCP server at https://api.entalpa.com/mcp. MCP, the Model Context Protocol, is a standard way for AI tools to read and update data. Connect an agent has the steps for each tool.
The first time, the agent opens a browser window where you sign in to Entalpa. From then on it works as you: it sees the projects you can open and changes the ones you can edit, and nothing else. These are the scopes it can ask for:
| Scope | What it is for |
|---|---|
openid, profile, email | Who you are: your account, your name, your email address. Every connection needs them. |
roles | The roles your Entalpa account has. |
offline_access | Staying signed in: the agent renews its access without asking you to sign in again. |
No scope adds to what your account can do, and none takes from it: a connection reaches every project your account reaches. There is no read-only connection.
What it can read
Everything in the projects you can open, through the 34 tools that only read: stakeholders, stories, requirements and their status, subsystems and interface contracts, open questions, the suggestions waiting for review, saved versions, the Spec score and the project's message history. It can export a project, as files or as a portable file. And it reads your own profile, with your credit balance.
What it can change
What a person can change in the app, through 46 tools that can write: stakeholders, stories, requirements, subsystems and interface contracts; a requirement's status as the work moves; answers to open questions; saved versions, undo and redo; imports.
20 tools can use AI, as the same actions do in the app, and then spend your credits: drafting stories, generating and analyzing requirements, detecting subsystems, assessing the spec, and more. The MCP tool reference marks each one.
Locks
A lock keeps Entalpa's AI from rewriting a requirement when it generates. It does not stop a direct edit: an agent, like a person, can still change a locked requirement.
What it can delete
9 tools can delete: a project, a stakeholder, a story, a requirement, a subsystem or an interface contract, and accepting suggestions, which can remove what they propose to remove. The server marks each of them as destructive, so the agent's app can ask you before it runs one. Whether it asks is that app's setting.
Inside a project, changes can be undone, deletions too: undo steps back through the project's last 49 changes, and a saved version restores the whole project. A deleted project cannot be brought back.
What stays your call
- AI proposals wait for you. Generate, Analyze, Audit backlog and Propagate stories save what they propose as suggestions and change nothing until those are accepted: by you, in the app, where the agent's link opens them; or by the agent, when you tell it which to accept. The other AI tools that propose, such as detecting subsystems or linking stakeholders to stories, hand their proposal to the agent and save nothing.
- A review where your app can show one. In an app that displays MCP Apps views, the agent can show a run's suggestions as a review in which you accept or decline each change yourself.
- Your account. Beyond reading your profile, no tool manages it: no buying credits, no plan or billing, no sharing, no members.