Orrery Workbench

The verifiable AI workbench for science and engineering.

Agents do the technical work and hand back a receipt you can reproduce: what ran, on which inputs, which checks passed. Sensitive data can stay on your machine with local models.

Part of Orrery. Same plans, no new subscription: free on your own hardware.

Capability status

What ships today, and what does not

as of Oct 2026
Inspection, not certification

Receipts record what was run and what was measured. They do not certify a result, a product or a process, and outputs require qualified human review.

01 · Coding with receipts Available now

Every task ends with a receipt, not a promise.

Orrery proposes each change as a diff you approve, runs your tests, and records the result. The receipt shows what changed, which tests passed, preview screenshots, the model used and what it cost.

  • Tests, preview, cost, model on every run
  • Client budgets: a hard spend cap per client or project
  • AI-use log: one log per client per month, for contracts that ask for AI disclosure
  • Tutor mode: the agent explains and checks your work instead of writing it for you

Budgets, receipts and the AI-use log for client work →

Task receiptIllustrative

Task receipt

#142 · Add CSV export to results page

Approved and applied
Tests
38 / 38
Files
4
Task cost
$0.84
Preview
2 shots
  • Test suite passed

    38 passed · 0 failed · 0 skipped

  • Diff approved

    +186 / −12 across 4 files

  • Model

    Recorded on the receipt, with where it ran

  • Client budget

    $12.40 of $50.00 used this month

Illustrative mock built for this page. The task and numbers are made up.

02 · Vision checks on previews In development

A local vision model looks at the preview, so you do not have to squint.

Orrery already captures preview screenshots on coding runs. We are building checks that hand those screenshots to a vision model running on your machine and record what it found: overlapping text, a missing element, a layout that breaks at phone width.

Findings go on the receipt next to the tests. They are a second pair of eyes, not a pass mark: a person still decides.

In development. Not in a released build yet, and the mock on the right is a design sketch, not a screenshot.

Preview checkDesign sketch
Preview · 390 px · local vision model
  • Finding 1 · button overlaps text

    Region marked on the screenshot, for a person to confirm

  • Model ran on this device

    Screenshot not sent to a hosted model

Design sketch of a planned feature. Not available yet.

03 · Reproducible analysis Early

Results pinned to the data, code, environment and seed that made them.

An analysis is only as good as your ability to run it again. Each notebook cell records hashes of its input data, its code, its environment and its random seed. A rerun check executes the cell again and compares outputs, and the whole chain exports as a receipt you can hand to a reviewer.

Early. Orrery's research notebook exists today; provenance receipts, the rerun check and export are in development.

Notebook cell · provenanceIllustrative
[7] fit = model.fit(df, seed=20261009)
    summary(fit)
Example provenance pins for cell 7
Pinned inputHashRerun
Data · trial_export.csvsha256 9f2c…a41ematch
Code · cell 7sha256 03be…77d0match
Environment · lockfilesha256 c1d8…2f93match
Seed20261009match
Outputs identical on rerunExport receipt ↓
Illustrative mock of a feature in development. Hashes and file names are made up.

04 · Engineering inspection with Vellum Engineering Early access

Inspection checks the agent cannot grade for itself.

Vellum Engineering runs manufacturing inspect and audit checks against your parts and assemblies, on your machine. Orrery calls those checks, then re-runs and verifies them itself before anything lands on the receipt, so an agent cannot report a pass it did not earn. From there it drafts a first-article inspection for a qualified inspector to review.

Early access. The Orrery side of the integration is built. AS9102 form export is planned, not shipped. About Vellum →

First-article inspection draftIllustrative

First-article inspection · draft

BRK-204 Mounting bracket REV C · draft, not released

5 checks re-run by Orrery
Characteristics
14
Within tolerance
13
Out of tolerance
1
Mass
0.218 kg
Verified rerun
5 / 5
Example characteristics from a first-article inspection draft
No.CharacteristicNominalToleranceMeasuredResult
1Hole Ø A6.40 mm±0.056.42 mmin tol
2Hole spacing48.00 mm±0.1047.96 mmin tol
3Flange thickness3.00 mm±0.053.09 mmout of tol
4Bend angle90.0°±0.590.2°in tol
Inspection, not certification

A draft for a qualified inspector to review and sign. No structural, material or process approval is implied.

Illustrative mock styled after Vellum Engineering's review screen. The part and values are made up.

05 · Sensitive data: local models only Local models available

For data that should never leave the room.

Medical and clinical research data, controlled engineering data: some work cannot go to a hosted model. Point Orrery at models you run yourself (Ollama, LM Studio, vLLM) and keep hosted providers switched off.

With local models, the model runs entirely on your machine: your prompts, code and files are not sent to any model provider. Supports research workflows; outputs require qualified human review.

Sign-in, update checks and opt-in crash reports still contact Ephemerent; they don't include your prompts, code or files.

Orrery is a research and engineering tool. It is not a medical device and does not make a regulated process compliant. Check your own obligations before using it with protected data.

Workspace data boundaryIllustrative

Workspace

cohort-analysis

Local only
Model
local · on this device
Hosted providers
off
Your provider keys
not used
Files
stay in the workspace
  • Recorded on each receipt

    Which model ran, and where

  • Human review required

    Outputs are drafts for a qualified reviewer

Illustrative mock. The workspace name is made up.

Pricing

Same Orrery plans. No new subscription.

Free

$0 your own hardware

  • The full agent on models you run yourself
  • Receipts, budgets and the AI-use log
  • Sensitive data never has to leave your machine

Pro · Max · Ultra

20M / 50M / 100M hosted credits a month

  • Hosted models included, with a hard stop at the limit
  • Your own provider keys
  • Everything in Free

Plan prices and details on the Orrery pricing page.

Bring one real piece of technical work.

Tell us what you need to verify: an analysis, a part, a codebase. Early-access users help decide what ships next.

Orrery records and re-runs checks; it does not certify compliance with any law, regulation or standard, and agent output is not guaranteed. Product mocks on this page are illustrative. See Terms, Privacy and Security.