Scope an AI Agent to One Blog Project
Editor note: I could not verify whether Roami restricts a sign-in to a single project at grant time. If it does, document it in the "Layers of control"...
19 October 2026 · 3 min read
Editor note: I could not verify whether Roami restricts a sign-in to a single project at grant time. If it does, document it in the "Layers of control" section; if not, the project id in each call is the scope.
Key takeaways
- If you run more than one site, always name the project in your prompt. Roami's tools take a project id.
- Reduce risk further by switching connectors and individual tools off when you don't need them.
- Claude only accesses what you grant during authentication. Deny permissions that look unnecessary.
- The strongest control stays server-side: agents can draft and edit, never publish.
Why does scope matter?
A person with a travel blog and a marketing blog does not want a draft about one landing in the other. The risk is not malice, it is a vague prompt. An agent that is asked to "improve my SEO posts" with two projects connected has to guess which one you meant.
What are the layers of control?
| Layer | What you control | Where |
|---|---|---|
| Prompt | Which project the agent works on | Say it every time: "In the Roami project, …" |
| Tool call | Roami tools accept a project id | list_projects first, then pass the id |
| Client | Connector on/off per conversation; individual tools on/off | Claude's + menu and "Search and tools" menu |
| Grant | What you approved at sign-in | Review requested permissions and deny extras |
| Server | What an agent can do at all | Draft and edit only; publishing blocked |
The client controls come from Claude's custom connector guidance. The Roami behaviors come from how its tools are built.
A safe session opener
Paste this at the start of any session:
List my Roami projects. Then work only in the project named "[name]".
Pass its id on every call. Before writing anything, tell me which project
and which post you are about to change. Save everything as drafts.
Common wrong-project mistakes
| Mistake | What happens | Prevention |
|---|---|---|
| "Update my posts about X" with two projects | Agent picks one | Name the project |
| Reusing a prompt from another project | Edits land in the wrong place | Include the project in saved prompts |
| Asking for a bulk fix without scope | Touches more than you expect | Use a dry run and a count |
| Cursor/Gemini using a default project | Drafts in an unexpected project | State the project in the first message |
Tools such as create_post accept an explicit project id, and some tools default to the connection's project. Always pass the id yourself.
What about write actions?
Turn off tools you won't need. Anthropic specifically warns that Research can invoke connector tools without further approval, so disable write-capable tools before using it. Also review tool approval requests, and choose "Allow always" only for tools you trust to run unsupervised.
Scope by task
| Task | Connector on? | Write tools? |
|---|---|---|
| Ask questions about your library | Yes | Off |
| Audit and preview fixes | Yes | Off until you approve |
| Draft a post | Yes | On for that conversation |
| Anything unrelated to your blog | Off | Off |
Frequently asked questions
Can an agent see all my projects? list_projects returns the projects your account can see.
What happens if it picks the wrong one? The result is a draft, which you can delete or move. Nothing goes live.
Does the publish gate apply across projects? Yes. Agents can't publish in any project.
Try it
Try Roami. Related: Revoking an AI Agent's Access, The Publish Gate.
Sources (checked 18 Oct 2026)
Next step: connect an agent, or audit your archive first.
Drafted with an agent in Roami, edited and published by Aaghran. This blog runs on Roami.