Your local AI.
Give it a home.

No. You select and review the material. Each receiver controls its own permissions. Private Folio originals and excluded passages stay out of Braid networking and public exports.

The exact edition you review and explicitly approve. The split-view studio lets you include or exclude source blocks and redact public excerpts beside the article preview. Sensitive-detail warnings run locally and remain advisory. Editing public content invalidates approval. The state banner distinguishes local changes, an approved draft awaiting export, and content matching the latest immutable export; it does not verify your hosted website. Folio exports HTML or a site ZIP for your own hosting. Scheduled streams prepare review drafts. Choose Editorial or Retro layouts, add local images and related editions, and check phone or print previews. Classic HTML is a separate compatibility choice; live Braid connection invitations require Modern HTML.

You may add an invitation to a publication. A reader requests a connection; you decide whether to accept. Rig-card exchange, fingerprint approval, and scope controls remain explicit. A public page never grants Braid authority.

Ollama models are discovered automatically. For vLLM, LM Studio, or another OpenAI-compatible local or LAN server, open Server settings, enter its address and port, and choose a model. No API key is needed unless that server requires one. Your existing inference server can keep running alongside its independent interface.

A capsule has five fields: title, mission, decisions, constraints, and next steps. New and imported capsules start excluded. To supply a capsule to local chat, select it in the Channel, allow that Channel’s context, and enable Include Channel context for the request. Editing clears the capsule’s inclusion. The request receipt lets you inspect the exact fields and revision supplied to the model. Changing included capsules or turning inclusion off starts fresh model context for that conversation; earlier chat turns stay visible.

Export/import moves a local .braidmemory file after your review. It does not automatically synchronize rigs, share a Channel note, publish through Folio, or grant an MCP client access. The file is readable text. Its checksum detects changes; it does not establish who wrote it. Capsule inclusion is selected context, not a guarantee that the model will follow it. Turning inclusion off affects future requests; it does not erase earlier exports or copies.

With an explicit MCP grant, a client can discover its authorized rigs and Channels, read specifically selected context, propose a note or handoff, and inspect its own requests. Proposals wait for human review. Approval creates a local note draft or stages a handoff; sending remains a separate human decision. The MCP connection grants no remote execution or Folio publishing authority.

You can cancel an approved unsent handoff, revoke a client, or acknowledge an uncertain result without changing its delivery evidence. Completed work remains in Review history. In earlier 2.0 hands-on checks, the scoped MCP tools were tested with Codex CLI 0.164.0-alpha.11.1 on Mac: authorized discovery, selected context, note and handoff proposals, and own-request status. Checks covered human draft edits, rejection without delivery, and refusal of a fresh connection using revoked credentials. These results apply to that client version and those workflows. Claude Code, Cursor and OpenClaw remain unqualified.

Braid 2.1 uses a separate profile. Existing stable and preview profiles are not automatically migrated. Existing identities, Channels, grants, and history are not automatically migrated. Keep a complete private backup of the profile you use. The ZIP includes setup, validation, and profile guidance.

The Braid 2.0 preview package remains available, alongside the 1.9.5 fallback below.

Braid-2.1-2.1.0.zip · 95,626,988 bytes. SHA-256:

12c5daf0050fe3e6b0b5f81c346bf370217c25c27e78ac7c3d1fb515de0ef2d9

The ZIP contains the Apple silicon app and its runtime, companion CLI, public source, Python launchers, licenses, and validation notes.

The bundled native app is for Apple silicon Macs and has been exercised on macOS 15.5. Its configured minimum is macOS 13; other macOS versions and Intel Macs have not been qualified for this build. The app is ad-hoc signed, not notarized.

Windows and Linux use the included Python launchers, requiring CPython 3.10–3.13 with pip and venv. First setup needs internet or a prepared offline dependency pack. These are not bundled native Windows/Linux apps. In the 2.1 prerelease, the founder confirmed Windows startup, local Channel chat with the model labeled “nemotron 3.5 bolt,” and fresh-chat capsule recall. These are human reports, not an independent performance benchmark or full Windows qualification. Exact Windows context receipts, opt-out, edit invalidation, file portability, and restart checks remain pending. Linux execution remains unqualified.

For the prior package, download Braid 1.9.5. Your local model server and model weights are separate from both packages.