gnomon-mcp

<!-- mcp-name: io.github.lihtness/gnomon-mcp -->

The pointer on a sundial that turns shadow into time.

A small MCP server for the boring-but-essential utilities every model needs: dates, calendars, arithmetic, unit conversion. Use it so your assistant stops "next-token guessing" math and date math.

Why

LLMs are bad at arithmetic and date math by default. They produce plausible answers that are often wrong by a small amount — exactly the kind of mistake that's hard to notice in a long response. gnomon-mcp exposes deterministic Python implementations through MCP so your model can compute instead of guess.

When an agent should reach for gnomon

Anywhere the next plausible token is not the right answer. Concretely:

• Math that matters — anything beyond trivial mental arithmetic, anything with a decimal point, anything that compounds. Call calc.
• "What day is it" / "how long until" / "how long since" — the model's training cutoff is not today. Call now for a snapshot; calendar with until/since/diff for elapsed time; parse for natural-language dates ("next thursday").
• Date arithmetic across month/year boundaries — adding 30 days, finding a quarter-end, counting business days. Models routinely off-by-one these. Call calendar with add / business_days.
• Unit conversion — call calc_convert. Never eyeball "kg → lb" or "°C → °F".
• Table-row workloads — when the same kind of computation needs to run on every row of a table, both batch tools (calendar, calc) take a list and return a list in order. One call, N results.

The rule of thumb: if you'd ask a colleague to "just double-check that number," call gnomon instead.

How this compares to other MCP servers

Time and math already have several MCP servers — the official Time reference (timezone-only), mcp-time and mcp-datetime (date formatting / timezone), calculator-server (math + units, no dates), and bundles like agent-utils-mcp (regex / hashing / JWT). gnomon's lane is narrower:

• Batch-first. calendar(ops) and calc(expressions) take lists; one tool call covers a whole table column instead of N calls.
• A real `now()`. One call returns 18 fields — ISO week, quarter, fiscal year, day-of-year, is_weekend, … — instead of just {iso, tz}.
• *Dates and math and units in one wiring.* No need to compose three separate servers.
• Natural-language dates baked in ("next thursday", "in 3 hours") without a separate NLP server.

If you only need timezone conversion, the official Time server is enough. If you want a broad utility bundle (regex, hashing, encoding, JWT), agent-utils-mcp is a better fit. gnomon is for the boring date-arithmetic-and-arithmetic core, batched.

Tools

Calendar

Two tools:

• `now(tz?)` — standalone. Returns a rich dict snapshot of the current moment. One call gets you everything about "right now".
• `calendar(ops)` — batch dispatcher. Each item picks its own op. Designed for table-row workloads (e.g. one call computes time-elapsed for every row).

`now(tz?)` returns:

{
  "iso": "2026-05-25T14:30:45+00:00",
  "date": "2026-05-25",
  "time": "14:30:45",
  "unix": 1779345045,
  "tz": "UTC",
  "year": 2026, "month": 5, "month_name": "May", "day": 25,
  "weekday": "Monday", "weekday_num": 0,          # 0=Monday
  "day_of_year": 145, "week_of_year": 22,         # ISO week
  "quarter": 2, "fiscal_year_us_gov": 2026,       # FY starts Oct 1
  "hour": 14, "minute": 30, "second": 45,
  "is_weekend": False,
}

`calendar(ops)` operations:

Units for diff/until/since: seconds, minutes, hours, days, weeks.

Example — compute several things in one call:

calendar([
  {"op": "until", "target": "2026-12-31", "unit": "days"},          # days left in year
  {"op": "since", "source": "2026-01-01", "unit": "days"},          # days elapsed in year
  {"op": "diff", "start": "2026-01-01", "end": "2026-12-31", "unit": "days"},
  {"op": "weekday", "date": "2026-05-25"},                           # "Monday"
  {"op": "add", "date": "2026-05-25", "n": 1, "unit": "months"},
  {"op": "parse", "natural": "next thursday", "tz": "America/Los_Angeles"},
])

Calculator

Examples:

calc(["2 + 3 * 4"])                  # [14]
calc(["sqrt(16)", "sin(pi/2)"])      # [4.0, 1.0]
calc(["mean([1, 2, 3, 4])"])         # [2.5]
calc(["sum(range(101))"])            # [5050]
calc(["(25 / 100) * 100"])           # [25.0]

Future tools (sketches)

The same logic — if the model is likely to bluff it, expose a deterministic version — points at several more primitives worth building. None of these are implemented yet; they are candidates, listed roughly in order of bang-for-buck:

1. Text measurement — count(text, unit) for chars / words / lines / sentences / LLM tokens. Agents constantly miscount "how long is this" and "will this fit in the context window."
2. Regex match / replace — regex_find(pattern, text) and regex_sub(pattern, repl, text). Models hallucinate which substrings match a regex; a real engine ends the argument.
3. Structured-data extraction — jq(path, json) / jsonpath(path, json). Reading values out of a nested blob by path, without typos.
4. Hashing & encoding — hash(text, algo) (sha256, md5, blake2), encode(text, scheme) / decode(text, scheme) (base64, hex, url, jwt-payload). All things models confidently invent wrong.
5. Decimal money math — money(expr) evaluated under Python's Decimal with explicit rounding. calc is float-based and quietly unsafe for currency.
6. Holiday-aware business days — extend calendar.business_days with a country (or calendar) parameter so US/UK/IN holidays are excluded. The current implementation only knows weekends.
7. Cron describe / next-fire — cron_describe("0 9 * * 1-5") → human English; cron_next(expr, n) → next N firing times. Models routinely misread cron fields.
8. Token counting for a target model — count_tokens(text, model) via tiktoken / Anthropic tokenizer. Lets an agent budget its own prompts and outputs instead of guessing.

If you want one of these, open an issue (or a PR — each is a small self-contained module that fits the existing tools/ layout).

Install

Recommended: no install — run on demand via uv:

uvx gnomon-mcp           # serves stdio MCP, ready for any client
uvx gnomon-mcp --demo    # call every tool once and print the results (no MCP client needed)

Or install globally:

pip install gnomon-mcp

Wire it into your agent

All recipes assume uvx gnomon-mcp. If you prefer a pinned install, swap the command for gnomon-mcp (with no uvx).

Claude Code

claude mcp add gnomon -- uvx gnomon-mcp

Or edit ~/.claude.json / a project .mcp.json:

{
  "mcpServers": {
    "gnomon": { "command": "uvx", "args": ["gnomon-mcp"] }
  }
}

Claude Desktop

claude_desktop_config.json:

{
  "mcpServers": {
    "gnomon": { "command": "uvx", "args": ["gnomon-mcp"] }
  }
}

Cursor

~/.cursor/mcp.json (or .cursor/mcp.json in a project):

{
  "mcpServers": {
    "gnomon": { "command": "uvx", "args": ["gnomon-mcp"] }
  }
}

Continue

~/.continue/config.yaml:

mcpServers:
  - name: gnomon
    command: uvx
    args: ["gnomon-mcp"]

Any other client (generic stdio)

Spawn uvx gnomon-mcp as a subprocess and speak MCP over stdin/stdout. That is the entire integration.

Hosted / remote (HTTP transport)

For team-shared instances or agents that can't spawn a local subprocess:

uvx gnomon-mcp --transport streamable-http --host 0.0.0.0 --port 8000
# also supported: --transport sse

Then point your MCP client at http://<host>:8000/mcp (or /sse for the SSE transport).

Tell your agent to actually use it

The MCP tool descriptions are intentionally terse to keep persistent context cost minimal (~150 tokens for all four tools). The richer "when to reach for gnomon" guidance lives in a Claude Code skill that loads on demand.

Option A — Claude Code plugin (one command, recommended)

The plugin wires both the MCP server and the skill in one shot. Inside Claude Code:

/plugin marketplace add lihtness/gnomon-mcp
/plugin install gnomon@gnomon-mcp

That registers gnomon as an MCP server (auto-starts via uvx) and installs the on-demand skill. Skill body loads only when the task triggers it — persistent context stays ~150 tokens for the four tool descriptions plus ~40 tokens for the skill's name + summary.

Option B — manual skill install (no plugin)

If you've already wired the MCP server with claude mcp add gnomon -- uvx gnomon-mcp and only want the skill:

mkdir -p ~/.claude/skills/gnomon
curl -fsSL https://raw.githubusercontent.com/lihtness/gnomon-mcp/main/skills/gnomon/SKILL.md \
  -o ~/.claude/skills/gnomon/SKILL.md

Option C — paste into your system prompt (non-Claude-Code agents)

For agents without skill support, paste this short version into your system prompt or CLAUDE.md:

You have gnomon: deterministic tools for dates and math. Use them instead
of guessing.

- `now` — current moment (your training cutoff isn't today).
- `calendar(ops)` — batch date math: diff/until/since/add/weekday/
  business_days/parse (natural language)/format.
- `calc(expressions)` — batch Python eval; math + statistics + common
  builtins pre-loaded.
- `calc_convert(value, from, to)` — unit conversion via Pint.

Both batch tools take a list and return a list. Prefer one batched call
over many small ones.

Development

git clone https://github.com/lihtness/gnomon-mcp
cd gnomon-mcp
pip install -e ".[dev]"
pytest

License

MIT