Each final video should have an identifiable project with its brand, footage, decisions and reviews. "My footage" is reusable; a project is the specific job.
Keep each video's files
To work locally, choose one EditorBrain folder per workspace, open it in Claude Code or Codex and ask it to prepare the project. New documents and media go in videos/<id>/: originals in assets/sources/, references in assets/references/, images in assets/images/, audio in assets/audio/, plans in plans/, reviews in reviews/ and deliveries in renders/. The editable, the samples and the temporaries go in .editorbrain-local/<id>/work/, previews/ and tmp/, respectively. If the ID matches a shared internal folder, the map uses .editorbrain-local/project-workspaces/<id>/ for those three private destinations. The agent follows the returned map and keeps existing explicit links.
Automatic preparation requires the next compatible version and keeps existing locations; 0.53.0 does not include the new command. Originals in another folder or disk stay there. Creating this structure does not download, move or delete files. The local location registry is private to the computer; checkpoints keep the hashes of the files needed to continue. See prepare destinations with the CLI.
With MCP, the agent can use the relative map and its own local tools without the CLI. A chat without disk access uses the service's assets and documents; MCP does not create or check folders on your computer. Storage preferences still apply, including not downloading a local copy when cloud is the right choice.
Create from the chat and upload from a phone
Ask for a "new edit" and name the brand if it is not clear. The agent can create the project with your saved context. If it does not have the recording, it returns the private link to upload one or more files. Then write "they're in" in the same chat to continue. Accessible attachments are used directly; you do not need to upload them again.
When the install exposes these commands, the equivalent flow is:
editorbrain projects create new-edit --brand BRAND_ID --style STYLE_ID --output-profile OUTPUT_ID
editorbrain projects intake --project new-edit
IDs are discovered on the account; they are not names the agent invents. create saves the project; intake checks footage and returns the link with its workspace. In MCP they are create_project and get_project_intake. A retry keeps the same ID and never replaces a different project.
The link requires signing in and having access to the project. Creating, looking up or uploading files does not approve a sample and does not start generation or a render. The state "footage present" means there are linked files; the agent still has to check that they are the right videos and analyze their content.
Find and follow projects
editorbrain brands list
editorbrain projects list --brand my-brand
editorbrain projects progress --project my-project
The board shows the recorded state of the work. It is not a signal that the agent is still running on your computer. Marking it finished requires the matching acceptance of the final version; generating a file is not enough.
Resume after you close the chat
Ask "resume yesterday's video about the product." The agent searches by name, brand and activity, clarifies ambiguous matches and recovers the checkpoint, reviews, corrections and existing jobs before it produces. It does not need to depend on the previous conversation.
In the CLI, projects find "product" helps find the ID and resume ID recovers context and locates the registered files. MCP uses list_projects and get_project_continuity. Absolute paths are stored only in the computer's registry; another device needs the media to be uploaded or available there. Saving a checkpoint does not upload videos and does not approve the final.
For several agents, use push --project ID, expected revisions and the batch turns. Before you repeat an action whose response was uncertain, check its ID or receipt. Continuity commands.
Ten recordings or five videos from scratch
First decide which files form each video. Ten files can be ten pieces or ten takes of one. Then create separate projects and group them in a batch.
editorbrain batches list
editorbrain batches create --file batch.json
editorbrain batches get BATCH_ID
The create schema is in create_editing_batch in the MCP reference. The projects must exist and be authorized. The batch keeps turns, review and budgets; it does not replace each project's preferences.
- An agent claims a turn before it starts that project; it does not duplicate work that is already claimed.
- If a piece is blocked by footage or permission, it can continue with other eligible ones.
- Pausing or canceling stops new work from starting; it does not guarantee that a local render or a generation already sent will stop.
- An expired turn requires checking what happened before you launch a generation again.
Work with other people
The owner invites an email and assigns scope. A brand editor works inside assigned brands; a project reviewer sees the versions shared for that project. They do not automatically get access to every brand or to billing.
editorbrain team access
editorbrain team members
editorbrain team invite --email colleague@example.com --kind brand_editor --scope my-brand
This last command sends a real invitation if the service is enabled. Use the confirmed email and the correct scope. For client reviewers use project_reviewer and the project ID.
The recipient opens the link, signs in with that email and verifies it. Before they accept they see the workspace and the assigned brand or project; the preview does not grant access. Accepting keeps their personal workspace. Each collaborator connects their own account to their agent: do not share your key with the team.
The invited editor can work from their agent or from My projects on the web: brief, script, direction, footage and samples. Shared reviews keeps the public thread with the client. The owner administers the brand and the general preferences. To use the integrated AI tools, the editor needs a positive monthly limit that the owner sets in Team, and the usual authorization for the job. Credits come from the shared balance; the limit renews on day 1 of each UTC month. team_spending_disabled means you can only review or you do not have a limit assigned yet; team_budget_exceeded means the job is over what that limit has left. Ask the owner to review it; topping up credits does not raise the limit and splitting jobs does not let you skip it.
Free includes 0 invited editors, Creator 1 and Studio 5, besides the owner. An editor and their pending invitations count once per email, even if they work on several brands. Client reviewers do not take a seat. The team_seat_limit error asks you to review seats from Team and the suggested plan; access is not removed automatically when you downgrade.
You can also share a private review link so the client can comment with their name and approve with their name and email without creating an account. editorbrain review list --project ID or MCP list_review_links retrieves those comments and approvals. The email is not verified; the approval refers only to the linked file and does not change the production review state.
Sharing a team review is explicit; unshared versions stay private. Removing access does not magically delete copies the recipient already downloaded.
Read every correction on a version
editorbrain team reviews --project my-project
editorbrain team comments --project my-project --revision UUID
editorbrain team comments --project my-project --revision UUID --after CURSOR
reviews pages versions; its historical comment summary is limited to 100 for the whole project. To review one version use comments and follow nextCursor until it is null. Each cursor belongs to that version and workspace. MCP offers team_reviews and team_review_comments with the same split. Internal production comments are not shown to clients.
On the web you can write a correction at the current instant. The draft is kept per account, workspace, project and version while you use the same browser session. "Refresh versions and access" looks for what is new without resending your comments. Approving applies only to that sample; pending corrections need to be resolved against a new version before you approve.
The owner changes brands and projects from Edit permissions, without removing the person or inviting them again. They can also do it from the CLI:
editorbrain team members
editorbrain team permissions --user USER_ID --file permissions.json --confirm
permissions.json contains the current revision and the complete access list:
{"expectedRevision": 1, "grants": [{"kind": "brand_editor", "scopeId": "my-brand"}]}
Copy that member's permissionsRevision into expectedRevision. If you receive TEAM_PERMISSIONS_CHANGED, read the access again and reconcile the changes. Saving invalidates earlier invitation links. When someone is left only as a reviewer, their agent connections are closed; becoming an editor again requires reconnecting. To remove everything use Remove access. Files are kept and old keys are not reactivated. There is no real-time co-editing; saving keeps revision control.
Scope limits
The plan, credits and storage are shared across brands in the same workspace. Identity and preferences stay separate by brand and project. External integrations and likes belong to the matching member; another member's credentials are not used.
Creating a team does not make a member an administrator of the service. Permission checks also apply in the API and MCP, not only by hiding buttons on the web.
Next in guidesAgency: clients, portals and budgets →