io.github.proscar87/oura-mcp

oura-mcp

English | 简体中文 | 한국어 | Español

The Oura v2 API as an MCP server. All 19 collections, six tools, no dependencies beyond the MCP SDK.

One local day of heart rate is 1,231 samples across 2 pages

A client that doesn't follow Oura's next_token returns 1,000 of them — 81%, looking complete, with nothing saying otherwise. Measured against the real API on 9 August 2026, one person, one ring, 24 hours.

The test to apply to any Oura MCP server, including this one: does it take next_token — or a cursor, or a limit — as a tool parameter? If it does, pagination is the model's job, and a model that forgets to ask again produces a confident answer off partial data. This server paginates to exhaustion before it returns, and tells you how many pages it took.

That's one of four ways Oura under-delivers without saying so. All four are measured below, and all four are corrected here.

Install

Download oura-mcp.mcpb from the latest release and double-click it. Claude Desktop does the rest — no terminal, no Python, no Node. It runs on Oura's official sample data out of the box, and every sample response says so, so nothing can pass for your own sleep.

Prefer the command line? uvx --from mcp-oura oura-mcp.

Prefer no Python on your machine at all?

docker run -i --rm -e OURA_SANDBOX=1 ghcr.io/proscar87/oura-mcp

-i is not optional: an MCP server speaks over stdin and stdout, not over a port. Without it the container has no stdin, the handshake never arrives, and the client reports a server that doesn't show up.

In claude.ai or ChatGPT, in a browser? Those connect by URL, so the server has to live somewhere public: deploy it as a Worker on your own Cloudflare account — free, about ten minutes, yours alone. See Remote: claude.ai and ChatGPT.

---

The problem, measured

Oura does not return errors when it can't give you what you asked for. It returns something different, shaped like a correct response. These are the four we found by measuring against the real API on 9 August 2026:

1. Skip the pagination and you get a fraction

{ "data": [ ... ], "next_token": "eyJ0eXAiOi..." }

If next_token comes back and you don't follow it, you receive the first page and nothing warns you. One local day of heartrate — one person, one ring, 24 hours — is 1,231 samples across 2 pages. A client that doesn't paginate gets 1,000 of 1,231: 81%, looking complete. A month is ~37,000.

2. Asking for a single day returned zero records

end_date does not behave the same across collections:

And on top of that, workout filters by UTC date while reporting day in local time: at -06:00, asking for July 16–18 returned records from the 15th and 16th — before the requested start.

Here the range is inclusive on both ends, always. Two extra days are requested on each side and then trimmed, which is correct whichever way a given collection behaves — and stays correct when Oura changes it.

3. latest=true is ignored where it doesn't apply

Only heartrate and ring_battery_level honor it. In the other seventeen Oura doesn't error: it returns the entire collection. You ask for the latest record, you get ten, and you believe it's one. Here it's rejected before the request goes out.

4. A field that doesn't exist is silently ignored

fields=does_not_exist returns the complete record — the projection never happens — and fields=score,does_not_exist applies the good one and drops the bad one without a word. Here, fields that never appeared are reported under ignored_fields.

The pattern is always the same: you ask for one thing, you get another, and nothing warns you. That's why this package would rather shout than quietly under-deliver.

Installation

Try it with no credentials

pip install mcp-oura
OURA_SANDBOX=1 oura-mcp --check

The sandbox is official — it's in Oura's OpenAPI spec, with 34 mirror routes — and serves synthetic data without authentication. 18 of the 19 collections work there: personal_info doesn't, which makes sense, since it's the one returning email, age, weight and height.

This is the right order: first you watch the server work and learn the shape of the data, then you go get credentials.

With your own data

Oura stopped issuing Personal Access Tokens in December 2025. Existing ones still work; new ones can't be created. So there are two paths:

a) OAuth2 — the one that works today. Register an application at cloud.ouraring.com/oauth/applications with the redirect http://localhost:9876/callback/ — the trailing slash is required, the portal rejects the other form with invalid_redirect_uri.

If you registered on developer.ouraring.com instead, your app belongs to Oura's newer portal, whose token endpoint is a different one. The legacy endpoint rejects those apps on every refresh — so the registration works exactly once, until the first access token expires, and then fails forever with nothing explaining why. This server tries the legacy endpoint and falls back to the new one automatically; nothing to configure either way.

export OURA_CLIENT_ID="…"
export OURA_CLIENT_SECRET="…"
oura-mcp --authorize             # opens the browser, waits for the callback
oura-mcp --authorize --manual    # headless machines: you paste the URL back

The token is stored in ~/.config/oura-mcp/credenciales.json with mode 600 — or in the system keychain if you happen to have keyring installed, which is not a dependency of this package — and refreshes itself. oura-mcp --forget erases it.

b) A personal token, if you already had one.

export OURA_PAT="your-token"
oura-mcp --check

--check is the self-check: it reports which credential you're using, which scopes were granted and how long the access has left, without returning the token or a single health value. It reports the token's length, never the token. Error messages get copied and pasted into chats and issues; they have no business carrying anything else.

Connecting it to Claude Code

With the package installed (pip install mcp-oura):

claude mcp add -s user oura --env OURA_SANDBOX=1 -- oura-mcp

Drop OURA_SANDBOX once you've run oura-mcp --authorize.

If you use uv, nothing needs to be installed permanently:

claude mcp add -s user oura --env OURA_SANDBOX=1 -- uvx --from mcp-oura oura-mcp

The --from is required because the distribution is named mcp-oura and the executable oura-mcp. (This needs uv; without it the command above fails with "command not found", and pip install is the path to take.)

As a Claude Code plugin:

claude plugin marketplace add proscar87/oura-mcp
claude plugin install oura@oura-mcp

Connecting it to Claude Desktop

One click: download oura-mcp.mcpb from the releases page and double-click it. Claude Desktop installs it — no terminal, no JSON, no Python. It ships with sample data turned on, so it works before you have any credential at all.

When you want your own data, just ask it for something: it opens Oura's authorization page through Claude, waits for the callback, and retries what you asked. No terminal. That works because MCP has a mode for precisely this — URL elicitation — and the client does the opening.

The one thing Oura still requires is that every application be registered, so you need a client ID and secret from cloud.ouraring.com/oauth/applications once. That's Oura's rule, not this server's. oura-mcp --authorize remains for terminal users and for clients that can't show a URL.

Or by hand, in ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "oura": {
      "command": "/full/path/to/oura-mcp",
      "env": { "OURA_SANDBOX": "1" }
    }
  }
}

which oura-mcp gives you the full path. Claude Desktop does not inherit your terminal's PATH, so a bare name there fails silently — one of the most common mistakes when configuring an MCP server.

Remote: claude.ai and ChatGPT

Experimental. Tested end to end on Cloudflare's real runtime with Oura's sample data, and its Oura login against a fake of Oura — but not yet against a real deployment, a real Oura account, or claude.ai and ChatGPT completing the connection. If you try it, an issue saying how it went helps.

claude.ai reaches a connector from Anthropic's cloud and ChatGPT from OpenAI's, not from your computer — so a server on your machine is invisible to both. ts/worker is the same server as a Cloudflare Worker that you deploy on your own account, for your own Oura account only. Nobody runs a shared instance, so your tokens and your data pass through no one else.

It adds three things the local server does not need: its own consent page, which shows which app is asking and where the result will go; a check, after Oura's login, that the account is the one you configured — anyone else is turned away, and with no owner configured nobody gets in; and a single place where Oura's single-use refresh token is renewed, one request at a time.

Deployment, step by step, and exactly what has and has not been verified: ts/worker/README.md.

The tools

Six, not nineteen. A server with one tool per collection forces the model to pick among 19 similar names before knowing what any of them contain. Here the collection is a parameter and the catalog is consulted when needed.

All six declare themselves read-only, and that isn't a promise: there is no POST, PUT or DELETE anywhere in the package, and a test reads the source to keep it that way.

oura_today is the only one that exists for convenience rather than for correctness: "how did I sleep?" needs today's two records plus enough history to know whether they are unusual, which was four round trips and four chances to stop early. It computes nothing — no average, no delta, no "your HRV is up 12%". The days come back raw and the comparison happens where the method can be cited. Across nine years of real data, three out of four changes between consecutive measurements fall inside the metric's own normal swing, so a percentage without that context manufactures a signal rather than reporting one. Its one parameter is days, from 1 to 30, defaulting to 7.

Examples

Each one was run against Oura's sample data (OURA_SANDBOX=1) on 8 October

1. Responses are abridged; every sandbox response also carries synthetic, saying the numbers aren't yours.

1. "Is it working?" → oura_check

{"mode": "sandbox", "data": "synthetic, from Oura, not yours", "oura_responds": true,
 "unavailable_in_sandbox": ["personal_info"],
 "next_step": "drop OURA_SANDBOX and set your own credential to see your data"}

2. "How did I sleep, and how ready am I today?" → oura_today

Last night's daily_sleep and today's daily_readiness, with the 7 days before them, raw. It says so in the response: "computed": "raw records only: this call computed no average, no delta and no trend."

3. "What can you see in my Oura data?" → oura_collections

All 19 collections, each with what it carries and its shape: "daily_sleep": {"shape": "date_range", "carries": "sleep score and its contributors"}, …

4. "Show me my sleep scores for the last three nights" → oura_query(collection="daily_sleep", start=…, end=…)

{"collection": "daily_sleep", "n": 3, "pages": 1,
 "data": [{"day": "2026-10-05", "score": 80, "contributors": {…}}, …],
 "discarded_out_of_range": "3 record(s) from the ±2-day safety margin were trimmed. This is normal and nothing you asked for is missing: …"}

pages counts what was followed to the end. The two extra days on each side are how the range stays inclusive in every collection, and the response says they were trimmed rather than leaving them unexplained.

5. "Did my sleep get better these last two weeks?" → oura_compare(metric="daily_sleep.score", a_start=…, a_end=…, b_start=…, b_end=…)

{"period_a": {"n": 14, "mean": 76.5}, "period_b": {"n": 14, "mean": 76.5},
 "difference": 0.0, "noise_band": 2.823, "verdict": "within_noise",
 "reading": "The difference (0) is inside what this metric does on its own over periods this long (±2.823). It is not evidence of a change — and with these days, a real change smaller than 2.823 would be missed more often than seen."}

6. "Do my steps affect my sleep the next night?" → oura_relate(x="daily_activity.steps", y="daily_sleep.score", start=…, end=…, lag=1)

{"lag": 1, "pairs": 118, "verdict": "cannot_tell",
 "reading": "These days give 4.852 effective pairs of day-to-day changes, and at least 20 are needed. …"}

Oura's sample data repeats a short pattern, which leaves fewer than five effective pairs in 118 days. Real data usually needs about three months.

oura_compare: the one calculation, with its method

"Did my HRV go up since I stopped drinking?" is the question people actually bring, and handing over two averages answers it wrongly: daily metrics swing on their own, and a good night tends to follow a good night, so a textbook comparison calls ordinary noise a change about a third of the time (measured: 34–38% at the autocorrelation these metrics typically have).

oura_compare takes a metric — collection.field, such as daily_readiness.score, sleep.average_hrv or daily_activity.steps — and two periods, a_start/a_end and b_start/b_end. It answers with the two means, the difference, and the band this metric moves in on its own over periods that long, measured from your own preceding 120 days and corrected for that day-to-day dependence. Then one of three verdicts:

• outside_noise — the difference is larger than the band. The level differs; not why, and not that it will last.
• within_noise — it isn't. This is not "no change": with those days, a real change smaller than the band would be missed more often than seen, and the response says so.
• cannot_tell — fewer than 7 days with a value in a period, or fewer than 60 days of history to measure the band from, or a metric that never varied — Oura's sample data is like that. No band is guessed.

The method was chosen by simulation, not before it: with the band estimated from your history and a t critical value, noise is called a change at most about one time in twenty at every autocorrelation tried, and the test suite holds it to that. Today is left out, because it is still accumulating. sleep uses each day's longest main sleep, and the response states that rule. Asking many comparisons and keeping the one that crosses finds a crossing by chance — the response says that too.

Asked whether something is trending, the server tells the model to compare the first half of the window with the second. A dedicated slope test was simulated against that and caught only a few points more, and only over 60–90 days (tools/trend_study.md), so there is no trend tool.

oura_relate: correlation, with the ways it lies taken out

"Do hard training days lower my readiness the next morning?" Correlating two columns answers it wrongly in two familiar ways: two metrics that both rise on weekends correlate without touching each other, and so do two that both drift over months. Measured on unrelated simulated metrics, a plain correlation called them related 41–93% of the time with a shared weekly rhythm, and 58% with a shared trend.

oura_relate takes two metrics, x and y, a range start/end, and a lag in days (0–7). It removes each weekday's usual level from each metric, compares only day-to-day changes between consecutive days, corrects the number of pairs for autocorrelation, and returns the correlation with its 95% interval and the same three verdicts. Every null simulated — weekly rhythms, random walks, shared trends, 30% of days missing — stays at about 5% (measured up to 5.2%).

• Lag is yours to choose, once. Oura files a night's sleep and the next morning's readiness under the day you woke up, so activity on a day meets the sleep that followed it at lag=1. first_pair shows which day met which, so a wrong alignment is visible. Trying several lags is several tests, and because the method works on changes, a real relation at one lag also shows up, reversed, next to it — the response says both.
• It needs about three months. Under 20 effective pairs it answers cannot_tell, and with gaps in the days it needs longer.
• It measures co-movement, not cause and not direction. Something else can move both, and the response says so every time.

oura_query parameters

And what the response tells you when something didn't come out clean: truncated with continue_from naming the last day reached, pagination_cycle if Oura repeats a token, ignored_fields, discarded_out_of_range, uneven_columns, empty when a query comes back empty, and large_response when what's returned is heavy enough to matter.

Four more say something happened that you'd otherwise never learn:

• synthetic — this is Oura's sample data, not yours. It rides on every response in sample mode, which is how the extension ships, so a model can't report made-up numbers as your sleep.
• rate_limited — Oura refused with a 429 and a retry got through. The data is complete; the warning is about the next query. Oura sends no rate-limit headers on successful responses, so being refused is the only signal there is that you're near the ceiling.
• fields_split — fields arrived as "day,score" instead of ["day","score"] and was split. No Oura field name contains a comma, so splitting is unambiguous — but reinterpreting your input silently would be the same sin this whole package is about.
• cached — the answer came from this session's memory instead of from Oura. Only ever for a range that closed before today, because a day that has ended cannot gain records; today is never held, since the ring syncs whenever it likes. An empty answer is never held either — nothing tells "no data" apart from "the ring hadn't synced yet", and freezing the second would turn a temporary gap into a permanent one. It lives in memory and dies with the process: no health data is ever written to disk.

That last one comes from measuring: 30 days of daily_activity is 252,000 characters, and 87% of it is a single field, met, a per-minute MET series. Asking for three columns with fields brings those same 30 days down to 5,000 characters — 99% less. The server doesn't trim on its own — that would be under-delivering — but it does say what's heavy and how to ask for less.

(Parameter names are stable, documented here, and the tool descriptions the model reads carry the same information. They were Spanish through 0.2.0; the rename to English landed in 0.3.0 and is recorded in the CHANGELOG as a breaking change.)

What this server does NOT do

It doesn't analyze beyond oura_compare and oura_relate. No trends, no anomaly detection, no advice — which is exactly where other servers place their value.

The reason: an average computed in here reaches the model as a number without its method. Across nine years of real data, three out of four changes between two consecutive measurements fall within the metric's own normal oscillation. A server that hands over "your HRV is up 12%" without saying how much that metric swings on its own isn't informing you: it's manufacturing a signal.

oura_compare and oura_relate exist because they answer that objection instead of ignoring it: the band comes with the number. Trends don't have a method here yet that survives the same simulations — a slope over autocorrelated days is the easiest signal of all to manufacture — so they aren't here.

Everything else, you get raw. The analysis belongs where the method can be cited — for instance with cotejo, which draws exactly that distinction for blood biomarkers.

The 19 collections

Daily summaries — daily_sleep, daily_readiness, daily_activity, daily_stress, daily_spo2, daily_resilience, daily_cardiovascular_age, vO2_max

The detail the scores hide — sleep (stages, HRV, temperature, latency), sleep_time, workout, session, rest_mode_period, tag, enhanced_tag

High resolution — heartrate, ring_battery_level

No range — personal_info, ring_configuration

Date-range collections use YYYY-MM-DD. heartrate and ring_battery_level use ISO 8601 with time.

Other Oura MCP servers

There are several as of August 2026, and it's worth being precise about the differences. benngermin/oura-mcp paginates properly, with a resumable cursor. daveremy/oura-mcp shipped the end_date fix the same week we did. davidmosiah/oura-mcp has the most complete MCP surface. Pagination no longer distinguishes anyone.

What does, as far as we could verify: workout's UTC skew isn't documented in any of them, nor is rejecting latest where Oura ignores it, nor warning about fields that were never applied. And none of them treats not analyzing as a stated position.

Privacy Policy

This section exists because the Claude connectors directory requires one. It is short because there is little to describe: the server runs on your machine and talks to a single service, the Oura API.

What is collected. Nothing, by us. The health data you request goes from the Oura API to your MCP client and passes through no server of ours, because there isn't one.

What is stored, and where. Only your credentials, and only on your machine:

No health data is written to disk, and that is the constraint the cache was designed around rather than a claim made after the fact. Answers for a day that has already closed are held in memory only, for the life of the process, and --forget clears them. Nothing about your sleep survives the server exiting.

If you deploy the remote Worker, the same holds with one difference: it runs on your Cloudflare account instead of your machine, so that is where your Oura tokens live — in a Durable Object — and the apps' tokens for it are kept in KV hashed, with what they carry encrypted. Health data is still never stored: it is fetched per request and held only in the Worker's memory. No one but you operates it. Details: ts/worker/README.md.

Who it is shared with. No one. The only outbound connection is to api.ouraring.com, with your token, to fetch what you asked for. Oura's use of your data is governed by their privacy policy, not by this one.

How long it is retained. Credentials, until you delete them: oura-mcp --forget, or by removing the file. Health data isn't retained at all — it lives in the response and that's it.

Diagnostics expose nothing. oura_check reports the token's length, never the token; the profile's field names, never their values. The token is wrapped in a type that won't print even in a stack trace.

Contact. Repository issues.

A note on language

The repository is in English: the code, its comments, the tests, and the internal documents (AGENTS.md, ROADMAP.md, CHANGELOG.md).

It was written in Spanish through 0.2.0. The tool parameters were renamed in 0.3.0 — a breaking change, recorded as one in the CHANGELOG — and the prose followed. Anything still in Spanish is a storage key that cannot be renamed without orphaning credentials someone already saved, and there is a test saying so by name.

License

MIT.

---

mcp-name: io.github.proscar87/oura-mcp