Skip to content

MCPtoAI v0.2.3

Choose a tag to compare

@github-actions github-actions released this 22 Oct 16:13
· 20 commits to main since this release
c29ee87

MCPtoAI 0.2.3 — Release notes

MCPtoAI 0.2.3 improves how your paired computer handles local tools, sensitive files, and approved file transfers. Features vary by operating system.

What's new

  • Skills on macOS and Windows: manage agent skills with preflight checks, access restrictions for skill files and data, and device-side policy enforcement. Managed Skills are not yet available on Linux.
  • Universal File Artifacts: send supported documents, images, audio, and video from a paired device to chat after explicit approval. Transfers use authenticated downloads, file-size and workspace checks, expiration limits, and cleanup. Chat can play supported MP4 videos. Returning a file path alone does not upload that file.
  • Protected locations: built-in file tools check sensitive directories and filenames, including application credentials, browser profiles, SSH keys, shell histories, and Windows AppData. Symlinks and Windows path aliases are resolved before file access.
  • Checks before approval: file paths, file type, extension, size, and permitted workspace are checked before a confirmation prompt appears. Approval cards identify file types and sizes and distinguish creating a file from overwriting one.
  • Clearer tool results: actions distinguish completed, failed, blocked, and denied. Statuses are communicated with text and icons, not color alone.
  • Safer approval details: the agent, not model-supplied tool arguments, determines file metadata shown in approval prompts for supported file operations.
  • Video controls: in-chat playback includes native controls and accessible seek buttons, using five-second steps for short clips and ten-second steps for longer clips.
  • Cleaner API keys: when you paste an API key into the Linux CLI or Device Agent CLI, control characters are removed. Keys previously saved with these characters are automatically repaired.
  • More reliable Linux AI connections: TLS fallback handling helps Linux connect to AI providers when the usual certificate verification path is unavailable.

Security and known limitations

  • Terminal commands are not sandboxed. If terminal access is enabled and a command is approved or otherwise permitted, it can access resources available to your operating-system account. The protected-file policy for built-in file tools does not sandbox run_command, run_in_project, background jobs, or external MCP tools. Review commands before approving them.
  • Keynote .key files are blocked. The .key extension is also used for private keys, so this version does not provide a Keynote-specific exception.
  • Linux wheel packages are not cryptographically signed. When installing a downloaded wheel, verify its published SHA-256 and SHA-512 checksums against a trusted release source. A checksum verifies the downloaded bytes against the reference; it is not a substitute for a trusted signature.
  • Managed Skills currently support macOS and Windows, not Linux. File transfer availability and supported file types may vary by platform.
  • Local permissions reduce unintended access, but they are not a substitute for operating-system account security and careful review of powerful tools.

Upgrading

  • macOS and Windows: when version 0.2.3 is offered in MCPtoAI Desktop, use Check for Updates, download the update, and choose Restart & Install. Reopen Desktop and confirm both the Desktop and Device Agent versions. Keep the device paired and review permissions after updating.
  • Linux CLI: python3 -m pip install --upgrade mcptoai. After upgrading, check the installed version with mcptoai --version.
  • Docker: docker pull ghcr.io/mcptoai/mcptoai:0.2.3. Keep your persistent configuration volume when replacing a container; verify the image digest from the trusted release information.

See SECURITY.md for reporting security issues and the Linux guide for platform-specific setup.

Integrity and verification

macOS (Developer ID signed and notarized)

  • MCPtoAI-0.2.3-arm64.dmg (162,729,553 bytes)
    SHA-256 5b160a60cef9b1065b3dfa5d7cf962eaed8583b6ba905e9b17d3bb3a1f880278
    SHA-512 2cecb5ee3af60e6f853dfebde260ef7175eed37f1747fd03ce545ccf316666602cde43c1a35bfdb537a7a2593decd9b691ba9374bc8260223eca9b6c1587c236
  • MCPtoAI-0.2.3-arm64.zip (162,296,503 bytes)
    SHA-256 d3eb5bd0270b1d5145bd4a0f9f6cc1eda5445462fe80d48e195245892ed85d5b
    SHA-512 b654cedaa8da41769ee6721b230721c7cef696cd9a849137fa414a529ea0e58028a56b09d89f1d6d37bc6531df66bb5ace4ff3779f7c9baf574e6f8a6f2f7876

Windows (Authenticode signed, publisher BKTY LTD)

  • MCPtoAI-0.2.3-x64-Setup.exe (138,744,264 bytes)
    SHA-256 0a5324ce8c960547938494b5d09fa9916511f4cce16f4d4dabf7c318904cd1b9
    SHA-512 3520fd6c11a7ae64134f2a2b8b713d8a979dc420c400c843fb50b0280d32deb5663d9709828386c84c49d0896bc7b81310b5eabea82d55c731a192f99ed716f5

Linux CLI (unsigned; verify checksums; PyPI shows the same SHA-256)

  • mcptoai-0.2.3-py3-none-any.whl
    SHA-256 1650cd43d73f495145b8e9d96ff9fca7a4d13793c85aefb76295e0ea5104e2e0
    SHA-512 f140fc5c1d68032dc980f65a6e8a853bf38262807c96872cb5da4e9d4ca86a9e6e3d223a58b4afb45abc7a99f90ab95c28743efed328532f0ccd06367621f02b
  • mcptoai-0.2.3.tar.gz
    SHA-256 07decd4763cdacc6a6b10dd3cf28ac5d3ec62a2841a980c1e599a810810c2cd5
    SHA-512 b8716c45de939f63a0210420218f73bb5007a99a08a9d70abfc4983c11cd207738ec6ecd37f47149ac9d8b5409450a87d4454775448236384abdcd16ae967c84

Docker (signed with Sigstore)
ghcr.io/mcptoai/mcptoai:0.2.3 digest sha256:481aaf9f931aa0dab3a119812dda43875f9747c36d96949ecb213b7a365118cd

cosign verify \
  --certificate-identity "https://github.com/mcptoai/MCPtoAI_Core/.github/workflows/publish-linux-docker.yml@refs/tags/docker-v0.2.3-r2" \
  --certificate-oidc-issuer "https://token.actions.githubusercontent.com" \
  ghcr.io/mcptoai/mcptoai@sha256:481aaf9f931aa0dab3a119812dda43875f9747c36d96949ecb213b7a365118cd