Local-first AI coding: what stays on your machine

The first value should not require source upload

A developer can scan a repository, fetch context, review a diff, and save personal learning without a cloud workspace.

Many teams need proof before they can approve a new GitHub app or source-ingestion service. Brief’s local MCP loop is designed for that order: install the connector, scan the checked-out repository, assemble a task packet, review the local diff, and keep the receipt on the machine.

Local-first does not mean the coding agent itself is always local. Cursor, Claude, Codex, or another model provider may have its own data path. Brief controls the context and evidence it handles; teams should evaluate the separate privacy terms of the agent they choose.

Derived sync shares proof, not silent source copies

Cloud-sync-derived mode sends bounded summaries and receipts while withholding raw source-derived bodies.

A team workspace becomes useful when members can see that repositories were scanned, context was fetched, reviews ran, and learning suggestions are waiting. Those signals do not require uploading every rule body, code excerpt, or private document.

In derived mode, Brief can sync scan summaries, context receipts, and user-authored learning suggestions. Approved team rules and memories can flow back into later packets. Each sync attempt reports whether it was skipped, accepted, or failed rather than silently pretending the team received it.

• Local scan and task selection remain on the machine
• Receipts identify scope and provenance without becoming a source dump
• User-authored learning suggestions can enter the team inbox
• Raw source-derived rule, skill, and memory bodies stay withheld

Richer source sync is a separate decision

A mode name is not permission to ingest everything; source scope, retention, visibility, and deletion still need explicit controls.

Brief defines team-cloud as the explicit richer privacy mode, but managed team-cloud source sync is a roadmap boundary rather than something the local loop depends on. Product copy should not blur the existence of a mode with a claim that every source-management control has shipped.

Teams considering richer sync should decide who can enable it, which repositories and paths are included, how long data remains, who can retrieve it, and how revocation propagates to derived knowledge. Until those controls are approved, local-only or derived sync provides a narrower adoption path.

Ask data-flow questions per system

Separate Brief, the coding agent, the source host, and connected work tools when reviewing privacy.

A single “does it store code?” question can hide several vendors and execution environments. Draw the path: where the repository is checked out, where the agent runs, what Brief scans locally, what derived evidence is sent, and what a connected service receives.

Brief publishes data controls and security pages for its own boundary. A sound review pairs those documents with the policies of the selected coding agent and source host, then tests the configured mode rather than relying on a sales diagram.