The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Temporal MCP listing page.
Your model knows calculus but not what day it is. Fix that.
temporal-mcp is a tiny Model Context Protocol
server that gives LLM agents a sense of time between turns. Two tools, a few
hundred lines, stdlib + mcp + platformdirs. That's the whole thing.
Open a fresh chat at 11 PM. The model says "good morning." Resume a conversation three weeks later. The model picks up mid-sentence like no time passed. Ask for "today's status." Get yesterday's status. Or last Tuesday's.
LLMs don't have wall clocks. They don't know when the last user message was, whether the calendar flipped, or whether this is a fresh thread or one resumed after a long gap. Most of the time this is harmless. Sometimes it makes your agent sound like it just woke up from cryosleep.
A persistent per-thread last-seen log, exposed as two MCP tools:
temporal_tick — call this once per user turn. Returns "it has been
14 minutes since the last message, no day rollover, timezone MDT" in a
format the model can actually read.temporal_peek — same thing, but doesn't advance state. For when you
want the gap without claiming a turn.That's it. Time exists. Your model should know that.
You'll see something like:
Run it twice and the second response shows the gap. Run it tomorrow and
you'll get day rollover: yes. That's the whole point.
Every tick returns a human-readable header and a JSON payload:
The header is for the model. The JSON is for your code, in case you want to
do something interesting with day_rollover (greet differently, reload
context, recompute "today's items") or with delta_sec (decay relevance,
detect a resumed session, flag idle threads).
If you use claude.ai web, ChatGPT, or anything else that wants a remote MCP server, point your connector at:
There are two ways to authenticate, depending on what your client UI exposes:
Both claude.ai and ChatGPT's custom connector UIs require OAuth 2.0 with
a Client ID and Client Secret. The hosted endpoint is a full OAuth
provider — visit https://temporal-mcp.dev/connect and click
Generate OAuth Credentials. You'll get a fresh client_id +
client_secret pair, shown once. Paste them into your client's
connector config. That's the entire signup.
No email, no password, no account record — the credential pair is the identity. We store only a SHA-256 of the secret, so we never see the plaintext. Generate a new pair any time you want a fresh timeline.
Claude.ai setup: Settings → Connectors → Add custom connector. URL
https://temporal-mcp.dev/mcp. Paste your Client ID and Client Secret.
Connect. The auto-approve flow redirects you back, claude.ai exchanges
the code for a token, and you're done.
ChatGPT setup: Same idea — Settings → Connectors → Custom MCP. Same URL, same credentials.
If your client supports custom HTTP headers (most do), skip OAuth and just send any opaque string as a bearer token:
Pick a UUID, a passphrase, anything. We SHA-256 it before storing anything; same identity-is-the-credential property as the OAuth flow, without the dance. This is the original lowest-ceremony path and works for any client that lets you set a custom header.
If your client's connector UI only exposes a URL field — no headers, no auth, no OAuth — embed your token directly in the path:
Or as a query parameter, if the path form gets stripped:
Same SHA-256 hashing, same identity model. URL-embedded tokens leak more easily than header tokens (proxy logs, referrers), so this path is a pragmatic fallback rather than the default — but the threat in our model is "someone advances your timeline," not data exposure. Rotate by picking a new random string any time you suspect the URL has been logged where it shouldn't be.
No signup. No email. No PII. The hosted endpoint is free, rate-limited to 60 requests/minute per credential. If you outgrow that, self-host (see below).
For desktop/IDE MCP clients, pip install the Python package and run it
locally. No network round-trip, state lives on your disk, no auth needed.
Python 3.9+. Linux, macOS, Windows.
Run as stdio:
Same idea — point the client at the temporal-mcp command.
The hosted endpoint at temporal-mcp.dev runs on Cloudflare Workers backed
by D1. If you want your own instance — for privacy, scale, or to ship it
as part of a larger product — the entire deploy lives in
workers/:
Free tier covers ~100k requests/day forever. Set
REQUIRE_AUTH=true in [vars] to refuse anonymous traffic. The Worker
is ~400 lines of TypeScript and has its own unit tests
(workers/test/).
temporal_tickAdvance the clock for a thread and return a snapshot. Call once per user turn.
| Field | Type | Notes |
|---|---|---|
thread_key | string, optional | Stable conversation/session ID. claude.ai web: conversation ID. Cursor: window/workspace ID. Anything else: any caller-stable string. Omit it and you get a default hostname+cwd hash — fine for local testing, not for serving multiple threads. |
client_id | string, optional | Namespace tag (e.g. "caweb", "cursor"). Defaults to "mcp". Use distinct tags per client so threads don't collide in shared state. |
temporal_peekRead-only. Same shape, doesn't advance state. Use it when you want the gap delta but the call isn't the canonical "one tick per user turn" event.
Per-thread last-seen state lives at:
| Platform | Path |
|---|---|
| Linux | ~/.local/share/temporal-mcp/state.json |
| macOS | ~/Library/Application Support/temporal-mcp/state.json |
| Windows | %LOCALAPPDATA%\temporal-mcp\state.json |
Override with TEMPORAL_MCP_STATE_DIR=/some/path.
State writes are flock-safe on POSIX and atomically replaced via
os.replace, so multiple agents pointing at the same state directory will
not corrupt each other. (Windows falls back to an in-process lock — fine
for a single MCP server, not designed for cross-process contention.)
Not exposed as an MCP tool on purpose — a model that can prune its own memory of "when did we last talk" will eventually do it at exactly the wrong moment. Run it from cron if you care.
Thread keying is namespaced as {client_id}:{key}. Reserve a unique
client_id per surface so threads from claude.ai web don't collide with
a local Cursor session sharing the same state directory.
Failure is honest. If the state file is unreadable or the lock times
out, the snapshot returns available: false with an error field and
the header says gap: unknown. It does not silently lie and call it
a fresh thread — a model that thinks every turn is fresh will keep
saying good morning forever.
Watchdog. tick() runs in a daemon thread with a 100 ms timeout so
a stalled state read can't block your hook budget. If it times out, you
get the honest-failure snapshot above.
No HTTP transport in 0.1. Stdio only — that's what Claude Desktop, Cursor, and the other major MCP clients actually use. HTTP/SSE can land in 0.2 if there's demand.
Mcp-Session-Id and friends, configurable timezone overrideresume: true flag past N
hours) so agents can branch on resumed sessions without doing the math
themselvesMIT. See LICENSE.
Built by Garret Sutherland / MirrorEthic LLC, extracted from the temporal layer of a larger cognitive-mesh project where this primitive was load-bearing enough to deserve its own package.
mcp-name: io.github.MirrorEthic/temporal-mcp