Use Private Mesh

Private Mesh is an opt-in core module managed from the top-level Mesh page. It shares existing local vLLM, Ollama and FreeToken backends with enrolled appliances and consume-only employee laptops. It never starts another copy of the model.

Core access and security

Private Mesh requires no license file on an appliance or consume-only companion. Administrators enable it through the authenticated dashboard API. Appliances and companions use signed invitations, device identities, membership, expiring signed rosters and per-model allow-lists. Consume-only companions cannot create meshes, invite members or export models. A license file never bypasses these checks.

Membership revocation, expired authorization leases and unavailable discovery fail closed. Reconciliation clears stale native allow-lists and owned remote routes/share keys. Local models and normal LiteLLM routes remain independent. Recovery actions include unsharing, revocation, leaving and disabling the module.

The implementation is in core/magicstick_core/private_mesh/, with BSL 1.1 terms in LICENSE. Upstream MeshLLM and endpoint-plugin code retain their own Apache-2.0 notices. See third-party obligations.

User workflow

1. Open Mesh → Enable Private Mesh as an administrator. The existing module/Flux lifecycle installs the optional component.
2. Choose Create Private Mesh, enter mesh/device names, select relay settings and choose ready local models to share. Sharing is off unless selected.
3. Under Invitations, create a single-use invitation for Magic Stick or Employee laptop. The token is displayed once; send it over a trusted channel. Invitations expire and can be revoked before use.
4. On another appliance, open Mesh → Join Mesh and enter the device name and the Magic Stick invitation token. The join entry stays visible after setup, including when disconnected. If the optional module is not installed, the dialog first offers Enable Private Mesh and waits for it to become available. If this appliance already belongs to a mesh, it first requires an explicit Leave current mesh confirmation; opening the dialog or canceling before confirmation never changes membership. One appliance belongs to one mesh at a time. Leaving stops sharing/imported routes, not local models; leaving on the owner also removes the other members' authorization when their leases expire. Use a fresh invitation to rejoin. Tokens are cleared when the dialog closes or joining succeeds and are never stored in browser storage. For a laptop, open the platform's companion application and use its Join form. No LiteLLM/Kubernetes credentials are given to the laptop.
5. Overview, Models and Network show status, imports, sharing limits and relay configuration. Buttons save only actual configuration changes.
6. Unsharing removes the allowed export immediately and then removes remote routes. Revoking a node denies new requests as soon as the exporter refreshes its signed policy; previously started inference is not retroactively canceled.

The companion opens a loopback browser UI itself; its authenticated API cannot configure infrastructure or publish models. The interface is titled Magicstick MeshLLM Endpoint. Its collapsible OpenAI-compatible API section shows the current loopback Base URL, a masked/copyable inference-only API key, exact model ID and curl example. The independent, per-launch Bearer key allows only GET /v1/models and POST /v1/chat/completions; it cannot read the browser session, join/leave the mesh, change relay settings or stop the client. Connection details require the authenticated browser session and are not embedded in the public HTML or routine status response. Host/origin checks remain in effect, and the listener is never opened to the LAN.

The desktop HTTP adapter currently supports non-streaming text chat only (stream: false); streaming requests are explicitly rejected. Its local port/key may change on restart. These desktop limits are distinct from the appliance-to-appliance Mesh bridge's streaming support. See client connection instructions. Its GitHub build workflow runs on relevant pushes to main, matching pull requests, or Run workflow in GitHub Actions. No application signing credential or appliance secret is required.

Permanent GitHub client downloads

Open the repository's Releases page and select a MagicStickMesh pre-release. Every complete, successful main client build publishes its four platform archives and SHA-256 files as release assets without automatic expiry. Versioned tags use mesh-client-<build>.<attempt>-<commit>; old releases and downloads are never automatically deleted or overwritten. Public release assets can be downloaded without a GitHub login.

Package	Target	Release asset
MagicStickMesh-macos-arm64	Apple Silicon / M-series, including M5; macOS 14 runner	.zip with MagicStickMesh.app
MagicStickMesh-macos-x64	Intel Mac; macOS 15 runner	.zip with MagicStickMesh.app
MagicStickMesh-linux-x64	x86-64 Linux; Ubuntu 24.04 runner	.tar.gz with the complete executable directory
MagicStickMesh-windows-x64	Windows x64; Windows Server 2022 runner	.zip with MagicStickMesh/MagicStickMesh.exe and the bundled runtime

Download the platform .zip or .tar.gz asset and extract it once. The GitHub-generated Source code archives are not client applications. Each macOS/Linux archive preserves executable permissions and symbolic links. All archives include onboarding instructions, source/build metadata and resolved Python/Cargo dependency inventories; the adjacent .sha256 file covers the complete platform archive. See the bundled client instructions.

The workflow also keeps its intermediate Actions → Artifacts downloads for 30 days. Those expiring copies are for CI/debugging; their expiry does not affect release assets. PR builds produce only these temporary artifacts, never releases.

To promote an older, still-unexpired successful main build without rebuilding, use Actions → Build Private Mesh companion → Run workflow on main and set source_run_id to that build's numeric run ID. Leave the field empty for a new build. Publication checks repository, branch, workflow, source history, all four platform results, archive checksums, packaged launch evidence and embedded build revision. Historical three-platform runs cannot be promoted with this four-platform workflow; build a new revision instead. Existing published releases remain unchanged. A draft becomes public only after all eight assets are uploaded and their remote digests match. Retries resume matching drafts; differing existing files are rejected.

GitHub may deny the CI token permission to create a tag for a historical commit with older workflow files. In that case, publication stops with the exact tag and source SHA. A maintainer can create that lightweight tag at the specified SHA using their normal repository access, then retry the workflow. Never point the tag at a different commit or broaden the workflow token's permissions to work around this check. Existing tags are verified and never moved.

Before upload, CI checks packaging contracts, host/target architecture, the exact patched native source, the bundled transport checksum/version, Windows native import policy and macOS signature integrity. It starts the actual frozen application on each native runner and checks its loopback UI, rejected unauthenticated access, authenticated status and shutdown using disposable state. This check neither opens a browser nor joins a mesh. A missing package or failed launch fails publication. These are test pre-releases: macOS apps are not Developer ID signed/notarized and Windows packages are not Authenticode signed. Real mesh enrollment/inference, complete SBOM review and platform release acceptance remain separate gates.

Windows companion package

Extract MagicStickMesh-windows-x64.zip and start MagicStickMesh/MagicStickMesh.exe from inside the extracted package. Keep the _internal directory beside the launcher. The package contains the native x64 transport and Python runtime; Python, Rust, Docker and a local GPU are not required. This build does not provide Windows ARM64 or 32-bit executables.

The native transport starts without an extra console window and receives only its isolated runtime directory and the required Windows system environment, not appliance/admin credentials. Membership is persisted under %LOCALAPPDATA%\MagicStickMesh. Quit endpoint stops the application; closing the browser alone does not. Windows may warn about the unsigned test download. Do not disable SmartScreen, Defender or other security controls; organization-approved signing and Windows client acceptance remain release gates.

macOS companion package

Build on the intended Mac architecture: Apple Silicon notebooks (M-series) use arm64. The existing build_companion.py bundles the native policy-enabled MeshLLM transport, Python runtime and dependencies in MagicStickMesh.app; notebooks do not need a separate Python or Rust installation. It uses only the consume-only client interface and still requires a Client / Employee laptop invitation from the mesh creator.

PyInstaller rewrites and ad-hoc signs the native executable on macOS. The builder therefore records the checksum of the packaged executable, then re-seals and verifies the outer app without signing nested executables again. The resulting test bundle is not Developer ID signed or notarized. Do not describe ad-hoc signature verification as Apple distribution approval. For a release, sign the nested code with the release identity, regenerate its checksum, sign the outer app and complete notarization and launch acceptance. Never disable Gatekeeper or remove application security controls as a packaging step.

The macOS cryptography binding must use statically linked OpenSSL. Intel builds may compile this pinned dependency from source; CI sets OPENSSL_STATIC=1, disables the pip wheel cache and rejects dynamic libssl/libcrypto imports before building the transport. This avoids bundling incompatible OpenSSL libraries from Python and Homebrew under the same file name. The final frozen launch check remains mandatory on both Mac architectures.