Skip to content

Releases

mcp-patchbay v0.1.1

Choose a tag to compare

@github-actions github-actions released this 16 Nov 04:45
5e0a847

Added

  • Zero-config first run. When the default config, ~/.config/mcp-patchbay/servers.json, does
    not exist, the patchbay creates it with no servers ({"version": 1, "servers": {}}; on POSIX
    the file at mode 600 and, if it has to create it, the directory at mode 700), logs one stderr
    line saying where and how to add servers, and serves its own tools. Before, it exited 78. An existing group- or
    world-writable directory is still refused before anything is written, and a missing
    --config or $MCP_PATCHBAY_CONFIG file anywhere else is still an error (exit 78).
  • A leading ~ or ${HOME} in --config and $MCP_PATCHBAY_CONFIG is expanded. An empty
    --config= is now a usage error (exit 64) instead of meaning the default.
  • MCP Registry listing as io.github.distronode-corporation/mcp-patchbay: server.json, the
    mcpName field in package.json, and a release step that publishes it with
    mcp-publisher through GitHub OIDC once the npm version is out.
  • A Claude Desktop extension bundle, mcp-patchbay-<version>.mcpb, built and validated with
    the MCPB CLI, smoke-tested on macOS, Windows and Linux, and attached to each GitHub release.
  • PRIVACY.md and a Privacy Policy section in the README: the patchbay collects and sends
    nothing.

Documentation

  • README Quick start: claude mcp add comes first and works on its own; adding connectors is
    the next step.

mcp-patchbay v0.1.0

Choose a tag to compare

@github-actions github-actions released this 14 Nov 16:05
0a5454b

Added

  • An MCP server over stdio that hosts configured MCP servers as child processes, with the tools
    servers, enable, disable, restart and info. Each carries
    _meta: {"anthropic/alwaysLoad": true}.
  • Policies allow, confirm and deny. confirm asks a person through a form-mode
    elicitation, and fails closed when the client has no elicitation support.
  • Live tool publishing with notifications/tools/list_changed. enable waits up to 2 s for the
    client's re-list.
  • Proxying:
    • pass-through tool definitions and results, with outputSchema not enforced;
    • cancellation and progress (original progress tokens);
    • roots/list and elicitation forwarding, and roots list_changed fan-out;
    • sampling refused;
    • merging of children's own list_changed;
    • a per-server timeoutMs.
  • Children:
    • full env inheritance with per-entry env;
    • a process group per child on POSIX;
    • tolerance of non-JSON stdout;
    • redacted stderr ring buffers;
    • crash isolation;
    • pidfiles, and reaping of leftovers after a crash.
  • Remote servers through mcp-remote 0.15.2:
    • header values kept out of argv;
    • MCP_REMOTE_CONFIG_DIR at 0700;
    • sign-in detection with a URL-mode elicitation (or a form fallback);
    • the state waiting-for-sign-in.
  • Autostart per project (CLAUDE_PROJECT_DIR), and instructions folded into the patchbay's own,
    capped at 4,096 characters.
  • Redaction of tokens, URL query strings, and entries' own env, args and headers in everything
    the patchbay emits.
  • CLI: import (from a Claude Code config), doctor, and rename-permissions.
  • The examples/write-gate PreToolUse hook example, with tests.
  • confirmVia, a top-level config setting: "elicitation" (the default, unchanged behaviour) or
    "host". With "host", enable and restart refuse to start a confirm server and point to
    a new tool, enable_confirm {name}. The host's own permission prompt for that tool is the
    person's approval. This makes confirm servers usable in clients that do not show form
    elicitation to the person, such as the Claude Code VS Code extension, which answers it
    decline. enable_confirm refuses allow and deny servers and is listed only in host mode.
  • import writes confirmVia: "host".
  • doctor prints confirmVia and warns about Claude Code allow rules that would approve
    enable_confirm without asking (the exact name, mcp__patchbay__*, mcp__patchbay, * and
    other wildcards) and about defaultMode: "bypassPermissions", in the user's settings and, with
    --project <dir>, the project's. In host mode a warning exits 1. --no-start skips starting
    servers.
  • rename-permissions never writes a rule that matches enable_confirm.

Fixed

  • Shutdown no longer orphans children that ignore SIGTERM. A detached helper sends SIGKILL to the
    remaining process groups after 500 ms, without delaying the patchbay's exit. Pidfiles are kept
    until a child's exit is observed, and the startup reaper now escalates to SIGKILL. After a
    crash, the leftover group members also get a SIGKILL follow-up.
  • A remote server's url is now treated as a secret. The whole url, each path segment, each
    query value and any userinfo (6+ characters) are redacted from info, lastError and logs.
    mcp-remote logs the url to stderr, and many endpoints carry their key in the path. The url
    is still on mcp-remote's argv (mcp-remote has no other way to take it), which is now
    documented.
  • Only mcp-remote children (url entries, an mcp-remote command, or npx mcp-remote) can
    trigger the sign-in prompt, and only while they are starting. The prompt is always labelled
    [<server>] . A re-auth request while a server is on is logged instead of prompting. The
    sign-in URL is no longer written to the log, not even in redacted form.
  • rename-permissions without --config now loads the default config ($MCP_PATCHBAY_CONFIG
    or ~/.config/mcp-patchbay/servers.json) and applies server prefixes. Before, deny rules for
    prefixed servers silently stopped matching. Without a config file, it warns that names were
    mapped as they are.
  • A restart racing a disable no longer turns a confirm server back on: the disable cancels
    the restart. An enable in that window joins the restart instead of showing a prompt whose
    answer was ignored. A crash marks the server failed at once (it used to wait 20 ms), so a
    restart right after a crash asks a person again.
  • Autostart names that are inherited object keys (constructor, toString, __proto__) are
    now rejected as unknown servers at load time (exit 78); before, they crashed initialize.
    The same hasOwn check now covers import --project and rename-permissions' prefix lookup.
    An env key named __proto__ is now kept as a key instead of being dropped.
  • A JSON syntax error in the config, in import --from <claude.json>, or in the
    rename-permissions settings file is reported as the file path plus line and column. Node's
    parser message quoted the surrounding text, which could include a token.
  • Kills can no longer hit an unrelated process through a reused pid. Post-exit cleanup and the
    startup reaper never fall back to signalling a bare pid. Pidfiles now record the process start
    time (Linux, v: 2), and the reaper requires it to match and requires the pid to still lead
    its own process group. A pidfile without a start time is not trusted on Linux.
  • The config permission check is now complete (POSIX). The file is opened once, fstated and
    read through that descriptor (no check-then-read race). A group- or world-writable (or
    foreign-owned) parent directory is refused. A symlinked config is accepted only if its target
    and the target's directory pass the same checks. A FIFO is refused without blocking.

Documentation

  • SECURITY-MODEL.md, SPEC.md and README now say exactly what happens to tool results: they are
    parsed with CallToolResultSchema, which drops unknown keys inside content blocks and
    rejects unknown content types. The docs no longer claim "unchanged".
  • The redaction guarantee now states its 6-character minimum and covers the url. The sign-in
    URL text matches the code.
  • README: npm test takes about 20 s, not 40 s.