temporal-mcp-server
temporal-mcp-server
MCP server for time, timezone, and duration tools.
Run it locally over stdio (Claude Desktop, Claude Code, any local MCP client), locally over HTTP, or use the hosted instance — same tools, same code, three ways to run it.
A public instance runs on Cloudflare Workers at https://time.somamcp.com/mcp:
claude mcp add --transport http temporal https://time.somamcp.com/mcpBuilt on somamcp (README · npm), which supplies the MCP plumbing, telemetry, and health/introspection endpoints for both runtimes. Time logic is pure and functional, using functype.
Tools
Tool | Purpose |
| Current time as epoch, UTC ISO-8601, and wall-clock in any IANA timezone |
| Render an ISO-8601 timestamp in a target timezone |
| Add or subtract an ISO-8601 duration, with calendar-aware month arithmetic |
| Elapsed time between two timestamps, in whole units plus a readable summary |
somamcp also registers an info tool and /health, /health/detail, /info, and /dashboard endpoints.
Behaviour worth knowing
Date units are calendar units; time units are exact. This is the distinction that makes DST come out right, and it follows ISO-8601 and Temporal:
Across US "fall back" | Result |
|
|
|
|
Both are correct, and they differ. "Tomorrow" is a calendar idea; "24 hours from now" is a physical one.
timezone governs the arithmetic, not just the rendering. Calendar units are applied to that zone's wall clock, so add_duration in America/New_York behaves the way a person in New York expects.
Month arithmetic clamps rather than overflows. P1M on 2026-01-31 returns 2026-02-28, not 2026-03-03.
Offsets are resolved per instant, not per zone. America/New_York reports -04:00 in August and -05:00 in January. DST comes from the runtime's tz database, so there is no offset table here to go stale.
Naive timestamps are read in the supplied timezone. 2026-11-01T00:00:00 with America/New_York means midnight in New York. It never falls back to the host's zone — that would answer differently on a laptop than on a Worker.
Timestamp parsing is strict. Only ISO-8601 forms are accepted; 17 Aug 2026 is rejected with a hint. new Date() would have taken it and resolved it against whatever zone the process happened to run in.
Elapsed components share one sign. time_between returns all of days/hours/minutes/seconds negative for a backward interval, so summing them is correct, plus a direction of past/future/same.
Errors carry a hint. An unknown timezone returns the bad value and the expected format, so a calling agent can correct itself instead of guessing again.
Running as a local MCP server
Stdio is the default and the mode local clients expect. Nothing is hosted, nothing listens on a port — your client launches the process and talks to it over stdin/stdout.
Claude Code
claude mcp add temporal -- npx -y temporal-mcp-serverClaude Desktop
Add to claude_desktop_config.json:
{
"mcpServers": {
"temporal": {
"command": "npx",
"args": ["-y", "temporal-mcp-server"]
}
}
}On macOS that file lives at ~/Library/Application Support/Claude/claude_desktop_config.json; on Windows, %APPDATA%\Claude\claude_desktop_config.json. Restart Claude Desktop after editing it.
Running from a clone
If you'd rather not go through npm:
pnpm install
pnpm build
pnpm start # stdioThen point your client at the built entry point:
claude mcp add temporal -- node /absolute/path/to/temporal-mcp-server/dist/node.js{
"mcpServers": {
"temporal": {
"command": "node",
"args": ["/absolute/path/to/temporal-mcp-server/dist/node.js"]
}
}
}The package also installs a temporal-mcp-server binary, so a global install (npm i -g temporal-mcp-server) lets you use that name directly as the command.
Working in this repo
A checked-in .mcp.json registers the local build as the temporal server, so Claude Code picks up your changes rather than the hosted instance:
{
"mcpServers": {
"temporal": {
"command": "node",
"args": ["dist/node.js"]
}
}
}Run pnpm build first — it points at dist/, so an unbuilt checkout has nothing to launch.
Verifying it works
The server speaks JSON-RPC on stdout, so you can drive it by hand:
printf '%s\n%s\n%s\n' \
'{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"smoke","version":"1.0.0"}}}' \
'{"jsonrpc":"2.0","method":"notifications/initialized"}' \
'{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"get_current_time","arguments":{"timezone":"Asia/Tokyo"}}}' \
| node dist/node.jsOnly JSON-RPC goes to stdout; logs go to stderr, so piping is safe.
Running locally over HTTP
For clients that speak streamable HTTP rather than stdio:
pnpm start:http # http://localhost:3333/mcp — override the port with PORTThis is the same server and the same tools; only the transport differs.
Running remotely on Cloudflare Workers
pnpm cf:dev # local workerd runtime
pnpm cf:deploy # build + edge-safety check + deploycf:deploy runs pnpm build first, which includes check:worker — so a bundle carrying a Node built-in fails before anything reaches Cloudflare.
Continuous deployment
Deploys run through Cloudflare Workers Builds rather than GitHub Actions, so no Cloudflare API token is stored in GitHub at all — Cloudflare connects to the repo through its own GitHub App.
Set it up once in the dashboard (Workers & Pages → temporal-mcp-server → Settings → Build):
Field | Value |
Deploy command |
|
Build command | (leave empty — |
Root directory | (repo root) |
Pointing the deploy command at a package script keeps the gating logic in version control; the dashboard holds one stable line. The Worker name in the dashboard must match name in wrangler.jsonc (temporal-mcp-server), or the build fails.
The build image ships pnpm and honours .nvmrc (ours pins Node 24). Non-production branches default to npx wrangler versions upload, so branch pushes produce preview versions without touching the live deployment.
The MCP endpoint is at /mcp. To require a bearer token:
wrangler secret put MCP_AUTH_TOKENWith MCP_AUTH_TOKEN set, unauthenticated calls to /mcp get a 401. Leave it unset and the endpoint is public — reasonable for a clock, not for much else.
Build provenance
scripts/deploy.mjs stamps the deploy with GIT_COMMIT, GIT_BRANCH, and BUILD_DATE, so the info tool and /info report exactly what is running:
curl -s https://time.somamcp.com/info # protected; also available via the `info` MCP toolWorkers Builds exposes WORKERS_CI_COMMIT_SHA and WORKERS_CI_BRANCH during the build, but build variables are not readable at runtime — they have to be forwarded as Worker vars, which is what the deploy script does. Running pnpm cf:deploy locally falls back to git rev-parse, and a deploy from a dirty tree is stamped <sha>-dirty rather than claiming to be a commit it isn't.
Connecting a client to the deployed worker
The public instance is served from a custom domain:
claude mcp add --transport http temporal https://time.somamcp.com/mcpWith a token set, pass it as a header:
claude mcp add --transport http temporal https://time.somamcp.com/mcp \
--header "Authorization: Bearer $MCP_AUTH_TOKEN"Health check: https://time.somamcp.com/health.
pnpm cf:dev serves the same thing on http://localhost:8787/mcp, so you can point a client at a local workerd instance before deploying.
Why the worker imports somamcp/edge
somamcp's root barrel re-exports helpers that import node:fs. Importing it from a Worker drags Node built-ins into the bundle. src/worker.ts therefore imports somamcp/edge, and pnpm check:worker fails the build if a node: import, a bare Node built-in, or the root somamcp specifier reaches the worker bundle.
See somamcp's edge runtime docs for the full story, including a feature parity table covering what does and does not work on edge — stdio, sessions, reportProgress, and multi-result resources are all Node-only.
The check walks the actual import graph from dist/worker.js rather than matching filenames — the bundler hoists code shared with the Node entry into a chunk with a generated name, and a filename glob would skip exactly the file most likely to carry a leak.
nodejs_compat is deliberately not enabled in wrangler.jsonc. If a Node built-in ever arrives, the build should fail loudly rather than be silently shimmed.
The alias block in wrangler.jsonc
xsschema (transitive, via fastmcp) probes for every schema library it supports — valibot, effect, sury — through dynamic import. We only use zod, so those branches never run, but esbuild still has to resolve the specifiers. They are aliased to an empty module instead of installing three unused libraries.
Architecture
src/
clock.ts pure time logic — Either<TemporalError, T>, no I/O, no globals
tools.ts MCP tool registration; takes a server, creates none
index.ts library surface (runtime-agnostic)
node.ts entry: somamcp -> stdio + httpStream
worker.ts entry: somamcp/edge -> export default { fetch }registerTemporalTools(server) takes the server rather than building one, so both entry points register identical tools. Nothing in clock.ts, tools.ts, or index.ts touches process, the filesystem, or any Node built-in.
Failures are values. Every fallible function in clock.ts returns Either<TemporalError, T>; the tool layer folds a Left into an MCP error result. Nothing depends on stack unwinding, which is what lets the same logic run unchanged on both runtimes.
Development
pnpm validate # format + lint + typecheck + test + build
pnpm test # 34 tests
pnpm check:worker # verify the worker bundle is edge-safetest/worker.spec.ts drives real Request objects through the Worker's fetch handler over the MCP wire protocol, so integration breakage surfaces in CI rather than after a deploy.
License
MIT