The 27 September AI Code Toolkit release renames the CLI and moves four MCP packages to SDK v2

2 hours ago

@agiflow/aicode-toolkitSubscribe

On the 27th of September, AI Code Toolkit shipped five packages at once. The command line tool, aicode-toolkit 2.0, renames the command you type and drops its own dependency on the MCP SDK. Alongside it, scaffold-mcp 2.0, architect-mcp 2.0, one-mcp 0.5 and style-system 0.2 moved to version 2 of the TypeScript MCP SDK, and the scaffold and architect servers fixed how they pick a port.

Ask about this presentation

Chapters

  1. 0:00AI Code Toolkit, 27 September
  2. 0:30The command is now aicode-toolkit
  3. 0:57MCP servers on SDK v2
  4. 1:27one-mcp 0.5.0 behavior
  5. 1:54Port 3000 fallback
  6. 2:21Dependencies and build
  7. 2:44Upgrading
Show transcript

  1. On the 27th of September, AI Code Toolkit shipped five packages at once. The command line tool, aicode-toolkit 2.0, renames the command you type and drops its own dependency on the MCP SDK. Alongside it, scaffold-mcp 2.0, architect-mcp 2.0, one-mcp 0.5 and style-system 0.2 moved to version 2 of the TypeScript MCP SDK, and the scaffold and architect servers fixed how they pick a port.

    Slide text
    AgiFlow/aicode-toolkit released 2026-09-27
    Agiflow

    The 27 September AI Code Toolkit release renames the CLI and moves four MCP packages to SDK v2

    aicode-toolkit 2.0.0 · scaffold-mcp 2.0.0 · architect-mcp 2.0.0
    one-mcp 0.5.0 · style-system 0.2.0

  2. Start with the command line tool, aicode-toolkit 2.0, and the change that can break a script. In 1.3, installing the package gave you a command called aicode. In 2.0 that command is called aicode-toolkit, and it runs the ES module build. The old name is removed, not kept as an alias, so if a hook or a script calls aicode init or aicode sync, rename it. The npx form from the README keeps working as written.

    Slide text
    aicode-toolkit 1.3.0 → 2.0.0 breaking · package.json "bin"
    1.3.0
    aicode
    → ./dist/cli.cjs
    2.0.0
    aicode-toolkit
    → ./dist/cli.mjs
    Unchanged: npx @agiflowai/aicode-toolkit init

  3. The sibling packages that shipped the same day carry the bigger change. scaffold-mcp 2.0, architect-mcp 2.0, the one-mcp 0.5 proxy and style-system 0.2 all moved to the split packages of the TypeScript MCP SDK, version 2.1. This is an SDK upgrade, not a new protocol. Clients still negotiate the protocol version, and stdio, Streamable HTTP and legacy SSE all stay available.

    Slide text
    AI Code Toolkit · 2026-09-27 changed · docs/MCP_SDK_V2.md

    Four MCP packages move to the TypeScript SDK v2

    scaffold-mcp 2.0.0 boilerplates and features
    architect-mcp 2.0.0 patterns and review
    one-mcp 0.5.0 the proxy
    style-system 0.2.0 themes and components
    @modelcontextprotocol/server · client · node · server-legacy 2.1.0
    Still served: stdio Streamable HTTP /mcp legacy SSE /sse

  4. If one-mcp is your proxy, version 0.5 brings two behavior changes, and strict capability errors stay as they were. Listing tools, resources or prompts now collects every page from a downstream server, so you may see definitions past the first page. And stdio clients now cap reads at 10 megabytes, so a very large image or response can close the connection. Unsupported-capability errors still fail the way they did in v1.

    Slide text
    AI Code Toolkit · 2026-09-27 changed · one-mcp 0.4.2 → 0.5.0

    one-mcp 0.5.0 lists every page and caps stdio reads at 10 MB

    All pages
    listTools, listResources and listPrompts collect every page from a downstream server
    10 MB limit
    a larger image or response on stdio can close the connection
    Unchanged
    unsupported-capability errors stay strict, as in v1 (enforceStrictCapabilities: true)

  5. One fix lands in scaffold-mcp 2.0 and architect-mcp 2.0, for HTTP and SSE mode. The old line passed MCP_PORT through Number. With the variable unset, that gave not-a-number, and the fallback operator let it through, because it only catches null and undefined. Now the code checks the variable first, so with no flag, config or variable, the server uses port 3000.

    Slide text
    AI Code Toolkit · 2026-09-27 fixed · scaffold-mcp 2.0.0, architect-mcp 2.0.0

    An unset MCP_PORT now falls back to port 3000

    scaffold-mcp · 1.3.0
    const port = options.port ?? fileConfig.port ?? Number(process.env.MCP_PORT) ?? 3000;
    scaffold-mcp · 2.0.0
    const port = options.port ?? fileConfig.port ?? (process.env.MCP_PORT ? Number(process.env.MCP_PORT) : 3000);
    Number(undefined) is NaN, and ?? only catches null and undefined.

  6. The rest is housekeeping you inherit. aicode-toolkit 2.0 moves to new major versions of commander, chalk, execa and js-yaml. Across the repository, Biome gave way to oxlint and oxfmt, the code compiles with TypeScript 6, releases are cut through GitHub Actions, and the dependency audit now fails on high or critical findings.

    Slide text
    AI Code Toolkit · 2026-09-27 changed · package.json, security.yml

    aicode-toolkit 2.0.0 moves to four new major dependencies

    aicode-toolkit 2.0.0

    commander 14 → 15
    chalk 5 → 6
    execa 9 → 10
    js-yaml 4 → 5

    repository

    Lint and format oxlint + oxfmt
    Compiler TS 5.9 → 6.0
    Releases GitHub Actions
    Audit fails on high, critical

  7. The command name is the change that can break a script. Rename aicode to aicode-toolkit wherever a script calls it, expect one-mcp 0.5 to list every page, and let scaffold-mcp and architect-mcp 2.0 fall back to port 3000. That is the 27 September AI Code Toolkit release: a renamed command line tool, and MCP servers on SDK v2.

    Slide text
    AI Code Toolkit · 2026-09-27 upgrade

    Rename aicode to aicode-toolkit when you upgrade

    aicode-toolkit 2.0.0: scripts call aicode-toolkit, not aicode
    one-mcp 0.5.0 lists every page; stdio reads cap at 10 MB
    scaffold-mcp and architect-mcp 2.0.0 fall back to port 3000
    Agiflow github.com/AgiFlow/aicode-toolkit

What happened in AI today, in your inbox

One short email each evening. Read it and you’re caught up.

Built in New York with ❤️