Security
How Blume limits what agents can do, where credentials are kept, what it can reach on GitHub, and the controls you have.
Blume is built around a few plain limits: agents get narrow keys, only a person approves, credentials stay on your computer, and conversation history is not uploaded. This page lists what Blume does today. It makes no claims beyond that.
- Local and cloud. Your imported conversations and the AI analysis of them stay on your machine. Delivery records, uploaded evidence and synced setup files are in the cloud. See What stays local and what goes to the cloud.
- Agent and person. Agents record work and report evidence. They cannot approve it.
- Read and write. Blume reads GitHub. It does not change anything there.
Everything an agent writes into a delivery becomes part of your workspace’s cloud record. Keep secrets out of it.
Agent keys and connections
- Each agent gets its own private key. Installing the plugin in Claude Code or Codex creates a key in your active workspace. It stays in that agent’s own settings on your computer.
- A terminal install command is shown once. It contains a key for your account and is not stored in the app.
- Each conversation is its own connection, tied to one project in one workspace. A connected agent can work on the deliveries of its project, and can read only its own inbox and the messages it sent.
- Other local agents hold no key. Agents added from More on the Plugin page point at Blume’s connection script on your computer, and the connection comes from the Blume app you are signed in to.
- You can revoke at any time under Plugin > Connections: a key, or one agent’s connection. Revoking a key cannot be undone.
The agent’s guidance tells it never to put a key in a task, a message or a generated file.
What agents cannot do
- Approve. Approving a revision, a merge or a release step is not among the plugin’s tools. A decision you give in chat is reported with your exact words and recorded as yours, with that origin shown.
- Manage connections. Creating and revoking keys is done by a person in the app.
- Treat another agent’s message as permission. Messages between agents are information.
- Grant itself rights through setup. Describing applications and release steps never grants an approval.
An approval is tied to the code it approved. A new revision, newly opened work or a failing check withdraws it. See Reviews.
The plugin guides the agent. It does not enforce rules inside the agent. What Blume controls is its own record: what counts as approved, and what the app shows as missing or stale.
GitHub access
- Read-only. Blume reads the repositories you choose, through the Blume GitHub App. It never merges, pushes, comments or approves.
- Only the repositories you pick. You can change the choice at any time.
- The authorization stays on your computer, encrypted in your system keychain. It is not sent to your Blume workspace in the cloud.
- What reaches the cloud is a pull request’s title, branches, state, checks and review decision, workflow runs and deployments. Its description, comments and changed files are read from GitHub when you open the code review.
Credentials on your computer
- The desktop app stores its account session and GitHub token locally, using operating-system encryption.
- For Claude Code usage on a Mac, Blume reads Claude Code’s credential from your macOS Keychain. macOS decides whether to allow that. See Make Claude usage work on Mac.
- For Cursor usage, you paste a session token into Blume. It is encrypted with your operating system’s keychain and only sent to cursor.com.
- Sign-in uses remote account and authorization records, and connected services keep their own.
Diagnostics
Blume sends aggregate usage counts and product diagnostics, such as usage events, app version, platform, timings and sanitized error details. They are designed to exclude conversation content and credentials.
The plugin can also report that its connection to Blume failed and later recovered. That report carries no account, workspace, conversation, repository, URL, error text or credentials.
Your controls
| You want to | Do this |
|---|---|
| Keep Blume out of a project | Set the project to Skip in Settings > Projects > Project choices |
| Read conversations only where agents use Blume | Set Show to Projects that use Blume in Settings > Projects |
| Cut off an agent | Revoke its key or its connection in Plugin > Connections |
| Stop Blume reading GitHub | Choose Disconnect in Settings > GitHub |
| Stop syncing a setup file | Remove the sync mapping on that device. Deleting the cloud item is a separate action. |
Deleting local Blume data does not delete cloud delivery records or uploaded evidence.
FAQ
Does Blume upload my source code?
Blume does not upload your repositories. Code reaches the cloud only when it is part of something submitted to a cloud feature: text an agent includes in a delivery update, a diff or report registered as evidence, or a setup file you chose to sync.