kairos-lang-mcp

An MCP server for Kairos, the schedule definition language (https://kairos-lang.org). It lets an AI client evaluate a schedule definition deterministically — every firing in a range, or the next N — validate a definition, and read the descriptor reference.

The server is a thin wrapper over the reference implementation's CLI: the tools return exactly what kairos list --json / kairos next --json print (the CliReport, schema at https://kairos-lang.org/schema/cli-report.schema.json). Points are epoch milliseconds and are the identity; dates are one-to-one civil labels. Out-of-coverage intervals come back as annotations, never as silently dropped points.

日本語: README.ja.md

Install

Node.js 20 or later. The server speaks MCP over stdio.

npx -y kairos-lang-mcp          # start on stdio (what an MCP client runs for you)
npx -y kairos-lang-mcp --version

Claude Code

claude mcp add kairos-lang -- npx -y kairos-lang-mcp

Claude Desktop (claude_desktop_config.json) and other stdio clients:

{
  "mcpServers": {
    "kairos-lang": { "command": "npx", "args": ["-y", "kairos-lang-mcp"] }
  }
}

Tools

tool	input	output
kairos_list	source, from, to, tz, supply?, explain?	CliReport for [from, to) — all firings, interval annotations, coverage summary
kairos_next	source, tz, from? (default: today in tz), n? (default 1), horizonYears? (default 10), supply?, explain?	CliReport for the next n firings; found < requested means fewer than n exist within the horizon (the CLI's exit code 2)
kairos_validate	source, tz, supply?	{ok, version, diagnostics: [{kind, message}], expressions} — lexing, syntax, static rules and premise resolution, without evaluating a range
kairos_reference	word, lang?	the reference page for one descriptor (roll, within, nth, segmentBy, …; README lists them)

Resources: kairos://llms.txt (the language overview for AI readers) and kairos://reference/{ja|en}/{word} (the same pages as kairos_reference).

tz is required

The CLI defaults to the machine's time zone; an AI caller has no machine, so every evaluating tool takes tz (an IANA name such as Asia/Tokyo). It decides how labels and the [from, to) endpoints are read. If the definition's premise declares tz:, evaluation uses that zone — tz still labels the output.

Recommended practice: define your premise once and keep it. A premise names the calendar system, the zone, the week start and the holiday tables:

premise JP {
  calendar-system: Gregorian; tz: "Asia/Tokyo"; wkst: Mon
  national = [2026-01-01, 2026-01-12, 2026-02-11, 2026-02-23, 2026-03-20, 2026-04-29, 2026-05-03..2026-05-06,
              2026-07-20, 2026-08-11, 2026-09-21..2026-09-23, 2026-10-12, 2026-11-03, 2026-11-23] covering: 2026..2026
  satSun = everyDay |> filter(d => weekday(d) == Sat or weekday(d) == Sun)
  bizDay = everyDay \ (satSun | national)
}

@JP
everyDay |> within(month) |> last |> roll(Preceding, on: bizDay)

Keeping that text next to your schedules turns tacit knowledge ("our business days") into explicit knowledge that every call reuses — and the covering: clause makes the table's expiry visible in the output (runwayDays) instead of letting it rot silently.

Supply

supply resolves external(kind: dates | instants) bindings with a static bundle, exactly like the CLI's --supply: { "binding" | "Premise.binding": { "dates": ["YYYY-MM-DD", …] | "instants": [ms, …], "covering": "2026-01-01..2026-12-31", "asof": "2026-10-01" } } — schema at https://kairos-lang.org/schema/supply.schema.json.

Errors

Errors are returned as data, not as protocol errors, so the model can read and fix them. kairos_list / kairos_next: {command, version, error: {kind: "usage" | "supply" | "static", message}} with isError: true. kairos_validate: {ok: false, version, diagnostics: [{kind, message}]} (a finding, not an error — no isError). kairos_reference: {error: {kind: "usage", message}, words} with isError: true for an unknown descriptor. usage is the shape of the call (bad zone, window over 10 years, source over 64 KB), supply a violation of the supply contract, static anything the language reports about the definition (lexing, syntax, static rules, evaluation). Diagnostics are in Japanese — the reference implementation's canonical output language; the report itself is language-neutral JSON (the same as the CLI's --json). The spec in English (https://kairos-lang.org/en/spec/) defines what each diagnostic means.

Limits

Evaluation window at most 10 years; n at most 1000; horizonYears at most 10; source at most 64 KB. These guard against accidents, not against adversaries.

Name

This package is kairos-lang-mcp — the MCP server for the schedule definition language at https://kairos-lang.org. It is unrelated to other npm packages that carry the name Kairos (such as kairos-mcp or kairos-mcp-server), which serve other projects that happen to share the word.

Versioning

The server's version tracks kairos-lang (same number, same contract); it depends on kairos-lang for evaluation and on @modelcontextprotocol/sdk for the protocol.

License

Apache-2.0. The bundled reference pages and llms.txt are copies from the same repository.