JARVISLabs

JARVIS Labs · Personal Intelligence. Engineered.

A real-world JARVIS.

Voice, memory, computer control, engineering and research, joined into one intelligence that runs on its owner’s machine and has to prove what it did.

Watch it work How it works

This is the real system. The footage was recorded on 22 September 2026 in one continuous 60 fps take. Every panel in it was opened by a typed request that JARVIS carried out itself.

Public record
Exported 22 Sept 2026
Capabilities registered
342
Live at the 17 Sept 2026 census
9
Tests collected, 17 Sept 2026
23,984
Cognition
Hosted today, local model in qualification

Watch

Ask. It acts. Then it shows its work.

Each clip is one request to the running system, what JARVIS did, and what came back, including the refusals. Recorded from the booted desktop, reviewed frame by frame for private information, silent, and captioned from the turn records.

JARVIS ran the thermal_bench case (run-8a6b69c66acd4dbe): efficiency 0.31358, 593.15 K, 376.296 W, 1.269 GPa, all labelled calculated. The whole turn took 5.1 s. ShowsA request reaches a real solver through the policy check, and the result is published with its provenance and stated assumptions. Does not showCalculated, not measured. The panel’s provenance run ID lagged the new run until a restart.
What happens, step by step
  1. 0:00 Asked: “Run the thermal bench simulation and show me the engineering results.”
  2. 0:07 Solver run published: provenance, solver, software revision, stated assumptions
  3. 0:16 The digital twin the readings belong to
Asked how Webb keeps its instruments cold, JARVIS read 3 of 3 sources from 3 independent origins in 19.9 s and marked every finding with the source it rests on. ShowsThe research desk retrieves real pages, keeps a capture of each, and ties every finding to a source. Does not showThe spoken summary was thin: it mostly repeated a source title. Source count is not a calibrated confidence.
What happens, step by step
  1. 0:00 Asked: “How does the James Webb Space Telescope keep its instruments cold?”
  2. 0:04 3 of 3 sources answered · the research panel opens
  3. 0:09 Each finding names its source · pages captured as read
Four typed turns: remember a fact, recall it, forget it, ask again. After the forget, JARVIS answers “I don’t know. You asked me to forget the earlier information, so I no longer retain it.” Captions quote the turn records. ShowsAn explicit forget removes the fact from what JARVIS will recall. Does not showThe exchange is spoken, so the words are in captions; the HUD’s conversation panel stayed empty during these turns.
What happens, step by step
  1. 0:00 “Remember that the website launch footage was recorded tonight on monitor two.” → “Noted, sir.”
  2. 0:06 “What do you remember about the website launch footage?” → “You told me it was recorded tonight on monitor two.”
  3. 0:11 “Forget what I told you about the website launch footage.” → “Forgotten, sir.”
  4. 0:17 Asked again → “I don’t know. You asked me to forget the earlier information, so I no longer retain it.”
A three-step research mission on Artemis news: 3 of 3 steps succeeded in 2.7 s of active time, and the dossier was saved as a durable, revisioned receipt. ShowsMissions run as durable steps with dependencies, a time budget and a saved receipt. Does not showA short mission. Long-running missions across restarts are not shown here.
What happens, step by step
  1. 0:00 Asked: “Start a research mission on the latest Artemis news, and show me the missions panel.”
  2. 0:04 3 of 3 steps succeeded in 2.7 s of active time · dossier saved as a durable receipt
JARVIS launched Calculator and verified it (“Calculator is open on monitor two”). Asked to type a sum into it, the planner reached for a tool it had not been granted, and JARVIS refused: “not a tool I have been granted”. ShowsDesktop launch and close are verified, and ungranted tools are refused rather than attempted. Does not showIt did not do the arithmetic. Operating inside the app was not reachable from conversation that evening.
What happens, step by step
  1. 0:00 Asked: “Open Calculator.”
  2. 0:03 Launched and verified: “Calculator is open on monitor two.”
  3. 0:08 Asked to type into it → refused: “not a tool I have been granted”
  4. 0:13 Closed on request
“Emergency stop.” took a 420 ms turn; the HUD went to STOPPED and stayed latched. JARVIS was shut down, restarted with --rearm, and came back HEALTHY 41 s after relaunch. ShowsThe E-stop is reachable by a plain request, and it stayed latched until an explicit re-arm at restart. Does not showNothing physical was running, so this does not show what the stop interrupted.
What happens, step by step
  1. 0:00 Asked: “Emergency stop.”
  2. 0:03 Stopped within a 420 ms turn · the latch holds until an explicit re-arm
  3. 0:11 Shut down cleanly, then restarted with --rearm
  4. 0:14 Back up: every consulted producer answered · HEALTHY
The owner’s view of governance: lifecycle stages, the approvals queue and the permissions matrix (13 licences × 5 uses, 52 of 65 permitted). ShowsThe HUD shows the policy posture and the approvals queue from the governance kernel. Does not showNothing was awaiting approval, and the lifecycle counters read zero for this session’s actions, so this shows the panel, not an approval being made.
What happens, step by step
  1. 0:00 The governance panel: action lifecycle stages, an empty approvals queue, 52 of 65 permissions granted
The engineering model the HUD binds to, rendered from its own geometry: twin/model.glb, 2,058 triangles. ShowsThe HUD renders the bound engineering model in 3D on request. Does not showA simple fixture model, not a production part.
What happens, step by step
  1. 0:00 Digital twin · twin/model.glb · 2,058 triangles
Screen recording of the HUD on 21 September 2026 with no backend running. Every readout says AWAITING TELEMETRY instead of showing made-up values. It shows how the HUD behaves when it has no data; it does not show any capability working.
Every recorded demonstration

Understand

One intelligence. A governed body.

The brain proposes; the body acts only through one fail-closed kernel. Every effect is read back from the real world and recorded as a receipt, and memory keeps where each fact came from.

See the full architecture
  1. Owner Voice, typed text, HUD, gestures
  2. Mind Cognition, context, capability discovery, memory, research
  3. Kernel E-Stop, policy, approvals. Every path, every time.
  4. Body Windows and browser, files, engineering and CAD, missions, agents
  5. World Read back independently. Verified, rejected or unverified; never assumed.
  6. Record Receipts, audit trail, health, and the truth ladder behind every claim.

Different

A chat assistant answers. JARVIS acts, then proves it.

Typical chat assistant JARVIS
Where it lives A browser tab on someone else’s servers. The owner’s own desktop. Reasoning uses a hosted model today; policy, memory and the E-Stop stay local.
What it remembers A conversation, until the session ends. Separate stores for episodes, facts, projects, procedures and preferences, each with provenance, confidence and validity dates.
What it can touch Text in a window. Windows applications, the browser, files, engineering tools and long-running missions.
How it acts Tool calls that are trusted as sent. Proposals that pass an E-Stop, a fail-closed policy engine and owner approvals, then get checked against the world.
How it reports progress “Done.” A receipt: verified, rejected or unverified, with the evidence and the date.

Capabilities

120 catalogued capabilities, each with its proof.

Every capability states how far it has climbed the truth ladder, from concept to release-qualified, and what it cannot do yet.

Workshop

From a spoken question to a solved model.

Units that refuse to mix, real solver runs, governed FreeCAD operations on STEP geometry, a revision-bound 3D scene, and slicer estimates. Every result says whether it is a calculation, a simulation, an estimate or a measurement.

  • Calculation computed from a model or formula
  • Simulation a solver run, not a physical test
  • Estimate a slicer’s time and material forecast
  • Measurement an instrument reading on the real machine
  • Physical result an inspected part. None yet: the print door refuses by design.
Enter the workshop
JARVIS ran the thermal_bench case (run-8a6b69c66acd4dbe): efficiency 0.31358, 593.15 K, 376.296 W, 1.269 GPa, all labelled calculated. The whole turn took 5.1 s.
What happens, step by step
  1. 0:00 Asked: “Run the thermal bench simulation and show me the engineering results.”
  2. 0:07 Solver run published: provenance, solver, software revision, stated assumptions
  3. 0:16 The digital twin the readings belong to
The engineering model the HUD binds to, rendered from its own geometry: twin/model.glb, 2,058 triangles.
What happens, step by step
  1. 0:00 Digital twin · twin/model.glb · 2,058 triangles

Evolution

From a voice prototype to a governed system.

60 dated milestones across three repositories, 3,272 commits on the production branch alone. Periods without evidence stay visible as gaps.

  1. 17 Apr 2024 First voice A browser prototype with a wake word, speech in and out, and a command table. 3 milestones
  2. 16 Jun 2026 The toolkit A multi-agent toolkit, a command HUD, memory consolidation and the first capability ledger. 7 milestones
  3. 2 Jul 2026 Formal phases Twelve build phases: runtime, policy, memory, voice, missions, recovery, vision, operator. 7 milestones
  4. 27 Jul 2026 Proof over claims A constitution and a truth ledger. The final gate was run honestly, and failed honestly. 9 milestones
  5. 1 Aug 2026 Space and forge Engineering solvers, spatial cognition, and the printer placed behind the approval path. 9 milestones
  6. 5 Sept 2026 Body and brain The ONE HUD, the approved voice, 44 body rows, and a local language model held to hard gates. 16 milestones
  7. 22 Sept 2026 Advanced Systems Spatial engineering workspace, CAD operations, frames and shadow roots, capability discovery. 9 milestones
Follow the whole story

Evidence

Measured, then measured again.

JARVIS Bench keeps the project’s own history of measurements, including the ones that got worse.

Tests passing in the full suite Then 9,949 tests 16 Aug 2026 Now 23,588 tests 17 Sept 2026 MEASURED Median conversation turn Then 3,083 ms Baseline, pre-optimisation Now 1,270 ms Plus task-relevant tool projection MEASURED HUD input to visible, p95 Then 40.2 ms 8 Aug 2026 Now 5 ms 13 Sept 2026 MEASURED

How much is proven today

Of 342 registered capabilities, 9 passed a live probe on the real machine at the last census. The rest are implemented, blocked or not yet built, and they say so.

  • 9 Live passed a live probe on the real machine
  • 1 Ready ready, awaiting a live probe
  • 248 Implemented code and tests exist; not live-proven
  • 16 Not probed not probed in this census
  • 47 Blocked blocked on a dependency or owner action
  • 21 Not built specified only
Registry census of 342 capabilities, recorded 17 September 2026 from docs/ledger/CAPABILITY_TRUTH.json. How the census works
Open JARVIS Bench

Building

What is being built now.

No dates are promised. Items move when evidence moves them.

Now

Next

Later

From the engineering journal

Participate

Shape what JARVIS learns next.

Suggest a capability, vote on other people’s ideas, and follow the build. Suggestions never give anyone access to JARVIS, and the owner decides what gets built.

Suggest an idea See the community queue

Most requested

  1. Loading the community’s ideas…