seconv: add seconv mcp, a Model Context Protocol server over stdio

[ivandrofly]

Summary

• New subcommand seconv mcp runs seconv as an MCP server over stdio, so AI clients (Claude Desktop/Code, Cursor, ...) can inspect and convert subtitles without a shell: {"mcpServers":{"seconv":{"command":"seconv","args":["mcp"]}}}.
• Seven tools, each a thin adapter over the Core helpers the CLI subcommands already use (SubtitleInfoGatherer, SubtitleLinter, SubtitleConverter, FixCommonErrorsRunner, ...), so behaviour cannot drift between the two entry points:
  • list_formats (optional substring filter), subtitle_info, read_subtitle (paged paragraphs of any supported format), lint_subtitle
  • convert_subtitle — format, output folder/name, encoding, overwrite, offset, fps/targetFps, renumber, duration/gap adjustments, delete first/last/contains, operations, FixCommonErrors / RemoveFormatting rule specs, container track numbers, OCR engine/language, timeCodesOnly, resolution
  • list_fix_common_errors_rules, list_remove_formatting_rules
• Stdout is reserved for JSON-RPC: Console.Out is redirected to stderr and the converter runs quiet; logs go to stderr (--verbose for debug level).
• Tool failures return isError with the real message (unknown operation, missing file, ...) rather than the SDK's generic placeholder; unknown operations are hard errors, not silent no-ops, matching the CLI's strict parsing.
• Result JSON uses the same field names as the corresponding subcommand's --json output.
• mcp is listed in --help, --help-json, the seconv README and docs/reference/command-line.md.
• Dependencies: ModelContextProtocol 2.2.0, Microsoft.Extensions.Hosting 10.0.11 (no trimming/AOT, so reflection-based tool discovery is unaffected by the single-file publish).

Test plan

• dotnet build src/seconv/SeConv.csproj — clean, no warnings
• dotnet test tests/seconv/SeConvTests.csproj — 393/393 pass, including 10 new McpToolsTest cases (one spawns dotnet seconv.dll mcp, lists tools via the MCP client SDK, calls subtitle_info on a fixture and on a missing file)
• Register in Claude Code with claude mcp add seconv -- <path-to>/seconv mcp, then ask it to describe an .srt and convert it to WebVTT with a 1 s offset — the output file appears and its time codes are shifted
• seconv --help and seconv --help-json list mcp
• Existing conversions and subcommands behave exactly as before (no change to any non-mcp code path)

🤖 Generated with Claude Code

https://claude.ai/code/session_01V1UEPy85BqcK2D9dHgfymB

[niksedk]

Thanks for the careful work here — it merges cleanly onto today's main, builds with zero warnings, the stdio test passes, and the stdout/stderr handling and error reporting are done right.

I'm going to close it anyway, because the intended AI surface for seconv is the CLI itself plus seconv --help-json and the --json subcommand output. Shell-capable agents (Claude Code, Codex, Cursor agent mode) read the schema once and then get every option; a hand-mirrored MCP tool set can only ever be a subset.

That subset has already drifted. convert_subtitle maps 28 of the 60 conversion options, and the three options added to the CLI since this PR opened (--ocr-prompt, --translate-prompt, --output-filename-append) aren't in it. Everything translate-related, the OCR URL/model/prompt options, --settings, PAC/EBU/ASSA/custom-format/multiple-replace files, teletext and plain-text options are all missing — and those are the parts where an agent most needs the full engine. Each future option would need a second parameter, description and test to stay in sync.

The dependency cost also matters for a single-file tool: ModelContextProtocol + Microsoft.Extensions.Hosting add 31 DLLs (~3 MB) and the whole config/DI/logging stack to run one stdio loop.

If MCP support for shell-less clients (Claude Desktop etc.) turns out to be wanted, I'd rather see it as a small separate project that generates its tool schema from CliSchema and exposes a single "run seconv with these args" tool, so there's one source of truth — but not as part of the seconv binary.