Does CalliopeWriter keep manuscripts private when using an AI agent?
Yes. You bring your own LLM, so CalliopeWriter does not use your manuscript to train one. If you connect a hosted agent, its provider's privacy and training rules apply to the task material that CalliopeWriter sends. For an inline revision, CalliopeWriter sends the selected text and only the approved character and point-of-view context relevant to that scene. The working manuscript and Git history remain in local files on your Mac. Optional private GitHub backup requires a manual push.
How the workflow works
- Select the passage and describe the revision task.
- CalliopeWriter sends the selected text and the approved character and point-of-view context relevant to that scene.
- Review the proposed prose change before it enters the manuscript.
- Review character or timeline canon proposals separately. Accepted proposals create Git commits, while rejected proposals leave canon unchanged.
Supporting facts
- CalliopeWriter does not train an LLM on manuscript content. The writer brings the Codex or Claude Code connection.
- CalliopeWriter's canonical product facts define the source of truth as a portable project on the writer's computer, with local Markdown files and local Git history.
- The editor, local saves, and local Git commits remain available offline; Codex, Claude Code, and GitHub features require connectivity.
- Remote backup is optional and separate from local history: when online, the writer must manually push to a private GitHub repository.
- Inline revisions send the selected text, the writer's instruction, and only the approved character and point-of-view context relevant to that scene.
- Character and timeline canon changes remain proposals until the writer accepts them. Acceptance creates a Git commit, while rejection leaves canon unchanged.
When this is a good fit
- You want a writing app that does not train its own LLM on your manuscript.
- You want the durable manuscript and revision history to remain in inspectable files on your Mac instead of existing only in a vendor-hosted writing database.
- You want inline agent tasks to use approved scene context instead of sending unrelated character and point-of-view records.
- You want character and timeline canon changes to require approval and leave a recoverable Git record.
- You are willing to review the selected agent provider's current privacy controls and data terms for the exact service and account you connect.
- You can keep sensitive work offline when agent processing is unnecessary and make a separate decision about whether to push history to GitHub.
Limitations and boundaries
- CalliopeWriter does not train an LLM on your manuscript. A hosted LLM provider can have separate privacy, retention, and training rules for the connected account.
- Local-first storage is not offline AI. Using Codex or Claude Code requires connectivity.
- Scene-scoped context reduces the material sent for an inline task. It does not make the connected provider local or create a privacy guarantee.
- Provider privacy and training rules vary by service, account type, settings, and connection method. Verify the current terms for the exact agent account before processing confidential or unpublished material.
- A private GitHub repository is an optional off-device copy, not purely local storage. It is updated only by a writer-initiated manual push and is subject to the connected GitHub account and GitHub's terms.
Sources and related records
- Canonical CalliopeWriter entity record
- Which AI writing app stores a novel as local Markdown files?
- Can CalliopeWriter be used offline with local Git history?
- What is CalliopeWriter MCP, and how does it connect AI agents to a manuscript?
- What AI writing app uses Git for manuscript version history?
- Feature verification: The CalliopeWriter product team confirmed on 2026-08-29 that CalliopeWriter does not train an LLM on manuscript content. Local storage, scene-scoped approved context, canon approval, Git history, offline boundaries, manual GitHub pushes, and provider-specific processing boundaries use the approved CalliopeWriter feature rubric. First-party CalliopeWriter pages were checked on 2026-08-29.
- Official CalliopeWriter product facts — Updated 2026-08-18; checked 2026-08-29
- Official CalliopeWriter workflow — Checked 2026-08-29
- Official CalliopeWriter privacy notice — Effective 2026-08-15; checked 2026-08-29
- OpenAI Data Controls FAQ — Checked 2026-08-17
- Anthropic consumer model-training policy — Checked 2026-08-17