Fully closed-loop, insulin-only (no meal announcement)

September 2026 update. CamAPS Liberty is now a commercially launched insulin-only fully closed-loop option in four European markets for T1D age 13+, excluding pregnancy. This research overview must no longer be read as a claim that every such system is investigational.⁴

Today's best automated insulin delivery (AID) systems are hybrid closed loops: the algorithm runs basal insulin and corrections, but the person still has to announce every meal — counting carbs and bolusing — because injected insulin is too slow to catch a meal on its own.⁵ The research frontier described here aims to remove that last manual step entirely: a fully closed-loop, insulin-only system where you eat without telling the device anything, and the algorithm detects the meal and doses for it.⁵ It is not one product but a path, pursued prominently by the University of Cambridge group (Hovorka and colleagues) alongside other academic and industry teams.⁵

Components. Like any AID system it is a CGM + insulin pump + control algorithm. What changes is the algorithm: it adds automatic meal detection and prandial dosing so no announcement is needed. Cambridge's version is the phone-based CamAPS HX app, the fully-automated sibling of the hybrid CamAPS FX.⁶

Trial outcomes. In a single-center crossover trial, 26 adults with type 1 diabetes and above-target control (mean HbA1c 9.2%) used CamAPS HX with ultra-rapid lispro for 8 weeks with no meal announcement. Time in range (3.9-10.0 mmol/L / 70-180 mg/dL) rose to 50.0% versus 36.2% on their own pump plus CGM (+13.2 percentage points), mean glucose fell from 12.0 to 10.7 mmol/L, and time below 3.9 mmol/L stayed low and unchanged (0.88% vs 0.64%) with no severe hypoglycemia or ketoacidosis.⁶ A two-center adolescent trial (CamAPS HX with Fiasp, 8 weeks, n=24) showed the same pattern: time in range 45.2% vs 32.3% (+12.9 points), lower mean glucose, and no rise in hypoglycemia.⁷ In adults with type 2 diabetes the effect was even larger — time in range 66.3% vs 32.3% and HbA1c 7.3% vs 8.7% — because the slower meal dynamics of type 2 are more forgiving of insulin's lag.⁸

Automation level. This is the whole point: no carb counting, no meal announcement, no pre-meal bolus.⁶⁷ In a supervised crossover study, UVA's fully automated RocketAP algorithm lifted time in range in the 6 hours after an unannounced meal to 83% (vs 53% with a legacy hybrid system), with no increase in hypoglycemia and improved overnight control — confirming that automatic prandial dosing can substantially blunt unannounced-meal excursions.⁹ Open-source AID communities pursue the same milestone using meal-detection and automatic micro-boluses (community practice outside the cited trials).⁵

How good is meal detection now? (July 2026.) A systematic review published in July 2026 pooled 69 studies of meal-announcement-free algorithms (screened from 1,205 papers) and gives the clearest picture yet. Median meal-detection sensitivity was 88% and precision 93% — the algorithms usually spot the meal, and usually aren't crying wolf. But they spot it late: typical detection latency was 25-40 minutes after the meal started, by which time glucose is already climbing. Reported time in range across the studies spanned 65-89%, a wide and heterogeneous range, and the review is careful about why: much of this work is still in silico (simulated), and the simulations reported more false positives than the real-world studies did. Read it as a frontier that is genuinely maturing, not as a set of numbers you should expect on your own body today.¹

The speed gap — the central limiter. The honest ceiling is postprandial: because subcutaneous insulin peaks ~1.5-2 hours after dosing, an unannounced meal spikes glucose before any reactive dose can act.¹⁰ This is why faster insulin matters more here than anywhere else, and a head-to-head trial directly tested it: faster aspart versus standard aspart in fully closed-loop with unannounced meals and exercise gave similar overall time in range (53% vs 58%, not significant), showing that today's "ultra-rapid" insulins are not yet fast enough to close the gap on their own.¹¹ Reviews point to inhaled or other ultra-fast routes, and adjuncts, as the missing piece.⁵

Exercise. Unannounced activity remains a leading unsolved problem — post-meal insulin-sensitivity shifts are hard to predict, so exercise is the event most likely to break true "set-and-forget" use.⁵

Ages, indications, access. Trial evidence spans adolescents (13+) and adults with type 1 (and type 2) diabetes.⁶⁷⁸ CamAPS HX provided research evidence for the commercially launched CamAPS Liberty option. Liberty is CE-marked for T1D aged 13+, excluding pregnancy, and launched in four European markets in September 2026. Other systems remain investigational; no US fully closed-loop indication is established in the sources reviewed here.⁵⁶

Is industry building this? Slowly, and not all of it for us. Medtronic's next-generation Vivera algorithm is designed around optional meal bolusing — an admission that the announcement burden is the thing worth killing — and is expected as a software update to the 780G/Flex; but it comes to us via a conference roundup rather than published data, and there are no type 1 outcome numbers for it yet.² Insulet presented a fully-closed-loop programme at ADA 2026 (the EVOLUTION 3 feasibility study, 64% time in range with no user settings) and has started the EVOLVE pivotal under an FDA IDE granted in March 2026, enrolling up to 350 adults, with a 510(k) filing targeted for 2027 and launch in 2028. Read the fine print: that programme is in insulin-treated type 2 diabetes, not type 1.³ It is a real milestone for fully closed-loop as an idea — type 2's slower meal dynamics are more forgiving of insulin's lag, which is exactly why it gets there first — but nobody should read it as a type 1 product on the way.

What's coming. Three threads converge: self-learning/adaptive algorithms that need no user tuning (a first-in-human Dexcom-based system lifted type 1 time in range from 38% to 56% with no meal announcement),¹² faster insulins to shrink the postprandial gap,¹⁰¹¹ and adjunctive therapies (amylin, GLP-1, or glucagon) to blunt the spikes insulin alone can't catch.⁵ The adjunct thread got an interesting data point at ADA 2026: a McGill study (26 adults with type 1) reported that fully closed-loop with adjunct pramlintide (an amylin analogue) did not meet its prespecified noninferiority criterion against hybrid closed-loop with full carb counting, despite similar point estimates — about 71% vs 72% time in range, with similarly low time below range — at the cost of nausea and infusion-site discomfort. In other words, similar point estimates alone do not establish that removing carb counting preserves glycemic performance. Two caveats: this is a conference report summarized by a community source, not published data we can check, and adding pramlintide means it is no longer insulin-only.² Together these threads define the insulin-only path to a true artificial pancreas — and the honest state of play in mid-2026 is that the algorithms have got good, and insulin is still the thing holding them back.