Orrery by Ephemerent
AI that shows
its work.
Orrery writes and tests your code, checks it in a real browser, remembers your projects across chats, and researches with sources it can quote. Every run ends with a receipt: what changed, what passed, and what it cost.
Free with models on your own hardware (Ollama, LM Studio, vLLM). Pro: $40/month for hosted models and any provider, cancel anytime. Windows, Linux (x64 and ARM64) and macOS (Apple silicon, preview).
works with
OllamaLM StudiovLLMDeepSeekGLMOpenRouterCodexClaude Code
Agent output can be wrong or unsafe; keep backups and review changes before running or shipping them.
Visible boundaryReversible workHuman acceptance
scroll ↓
MEET SENTINELA decision layer
A decision layer
for the work ahead.
A small model that scores the candidate next steps, so the agent takes the one most likely to work. Explore the decision trail in a live, interactive mission.
01Restore service↗
02Find the fix↗
03Verify before shipping↗
Real model responses. Clearly labeled simulation.
01openpick a folder
→
02askin plain words
→
03watchit edits, tests, checks
→
04decidekeep, revise, or undo
Every edit is a reviewable diff · tests run before it reports · keep your own backups
See it work
Build. Remember. Research.
Real runs in the Orrery desktop app: one on a local model on my own GPU (the free tier), three on DeepSeek V4 Flash through Doubleword. No one touched the code or the evidence: the agent wrote every file, ran its own tests and checked the page in a real browser, and every research claim is quoted from its source.
Free, on your own GPU
“Make a small flappy bird clone I can play in the browser.” A 27B model served by vLLM on my own machine writes the game, serves it, checks it in a headless browser and looks at the screenshot, then you play it. 16 real minutes, sped up 30×, $0 in model costs.
From an empty folder
One prompt builds “Halcyon Observatory”: an animated orrery, show listings and a ticket selector. 14 tool calls, tests passing, browser-checked, in 63 seconds for an estimated $0.009.
Then it remembers
It builds “Tidewave”, then a brand-new chat with no history picks the site up from workspace memory and adds a theme toggle and a live buoy panel. Two runs, tests passing, an estimated $0.026 in total.
And researches with sources you can check
“Does intermittent fasting lead to more weight loss than daily calorie restriction?” The Medicine field searches PubMed, ClinicalTrials.gov, medRxiv and more; deep research keeps 8 claims it could quote and verify, passes the evidence gate, and one click hands the run to the chat, which answers with citations. $0.03 of a $0.50 cap.
Real recordings of Orrery, edited for length; sped-up segments are labeled in each video. Costs are list-price estimates from each run’s token counts, not billed amounts.
Research
Answers you can check, not just read.
Ask a real question. Orrery searches the scholarly indexes for your field, keeps only claims it can quote from a source it found, and tells you plainly what the evidence does not settle.
01
Every field, 14 free indexes
Pick a field (medicine, physics, mathematics, education, law, the humanities and more) and Orrery searches the right sources: PubMed/Europe PMC, ClinicalTrials.gov, arXiv, INSPIRE-HEP, zbMATH, ERIC, DOAJ, HAL, OSF preprints, DataCite, OpenAlex, Crossref and others.
02
Quoted, checked claims
Deep research extracts findings with a verbatim quote, confirms the paper exists, and judges whether the quote supports the claim. The written report cites only verified claims.
03
An evidence gate
Each report shows whether its evidence passes: verified claims, contradictions, and the checks your field needs, such as trial registration and risk of bias for medicine.
04
A budget you set
Every run has a spending cap and a receipt. A typical deep run costs a few cents of its $0.50 default cap.
05
Connected to your work
Send any paper or report to the chat in one click, or ask the chat to run the research itself. Answers link back to the run, and findings save to your Notebook.
06
Honest about gaps
When the literature does not answer the question, the report says so instead of filling the gap. A run the model could not finish is marked as failed, not as done.
How it works
You ask. It works in the open. You decide.
Orrery works in a project folder on your computer. Describe the task and it plans, edits files, and runs your tests, with every step visible as it happens. Nothing is final until you accept it.
01Open a folderAny project, or an empty folder for something new.
02Describe the taskIn plain words, like you would to a colleague.
03Watch it workPlans, file edits, tools, and test output stream live.
04It checks itselfRuns your tests, and opens web pages in a real browser, before it reports.
05You decideRead the diff and the receipt: what changed, what passed, what it cost.
Built for real projects
Everything it does stays in view.
Plans, edits, test output, and costs in one place. The record helps you check the work; it does not replace your own review and backups.
01
Reviewable edits
Every change arrives as a diff, with a checkpoint after each applied step, so you can see exactly what moved and rewind a step.
02
Tests and browser checks
After editing it runs your tests and fixes what it broke, and it checks web pages in a real browser before it says it is done.
03
Remembers your project
Workspace memory carries decisions from one chat to the next, and you can see and edit what it keeps.
04
Knows your codebase
Ask about the project and it searches code by meaning, not just keywords, with @codebase.
05
Works with your tools
Use MCP servers and local tools, and connect other editors and agents through open, standard protocols.
06
Costs you can see
Every run shows the model, usage, and cost. Effort presets and budgets keep routine work cheap and deeper work deliberate.
Safety boundaries
Automation with a review boundary.
Orrery can make mistakes, and generated code may be insecure, incomplete, or unsuitable for your project. The product is designed to keep work reviewable: scope checks, permission gates, model-route labels, proof summaries, and usage receipts stay visible where supported.
Review before shippingYou are responsible for accepting, running, and deploying generated code.
No guaranteed outcomePerformance depends on task, context, models, tools, and provider availability.
Keep backupsUse source control and inspect changes before relying on them.
Cloud variesHosted routes, quotas, latency, and providers may change or pause.
Data and account boundaries
Clear about what crosses the network.
Hosted routes send the prompt and code context needed for a completion through Orrery Cloud to the configured provider, and Ephemerent keeps no relay content beyond fulfilling the request. Hosted-model requests are relayed to fulfill the request, not captured for training.
01
Explicit hosted route
Nexus labels the selected hosted route before a run starts.
02
Purpose-limited relay
Only the prompt and context needed for the completion are relayed to the configured provider.
03
Content is not metering
Stored usage counters are token totals and timestamps, not copies of prompt content.
04
Your review remains final
Keep source-control backups and inspect changes before accepting, running, or shipping output.
FOR DEVELOPERSYour application.
Your application.
The Orrery API.
Bring hosted-model generation and Sentinel decisions into your own tools. Start with familiar chat completions, explore structured decision endpoints, and manage keys and usage in one console.
01Create an API key↗
02Copy your first request↗
03Inspect usage and results↗
cURL, Python and JavaScript examples.
Pricing
Free on your own hardware. Pro for everything else.
Run the whole agent free on a model server you run yourself. Pro adds hosted models with credits included and any provider with your own key for $40/month, cancel anytime.
Free
$0
no subscription, no API key
The full agent (diffs you approve, rewind, tests, receipts) on Ollama, LM Studio, vLLM or llama.cpp, on this computer or your local network. Nothing is metered, because nothing runs on our servers. A free account is only needed to download.
Download freeFounding 100: the first 100 developers get a month of Pro free, no card, plus feedback on their workflow straight from the founder. Claim a seat →
Agencies and freelancers: per-client budgets, client receipts and an AI-use log for contracts that require disclosure. Orrery for agencies →
Hosted models
Pro
$40
per month, cancel anytime
Individual agent work on hosted models (DeepSeek, GLM and others, served via Doubleword through Orrery Relay) with credits included and Nexus operations.
- Hosted models - 20M credits/mo (about 107M tokens at list rate)
- Credits are vendor-cost units, so nothing is marked up and nothing is inflated
- Codex, Claude Code, and other providers with your own key: OpenRouter, OpenAI, Anthropic, Gemini, DeepSeek
- Nexus, Discord/mobile oversight, proof receipts
Max
$100
per month
Bigger pools for daily multi-agent work; the same routes, more headroom.
- Hosted models - 50M credits/mo (about 267M tokens)
- 2.5x the Pro pool
- More cloud/background run capacity
Ultra
$200
per month
The largest pools plus cloud automation capacity for research runs.
- Hosted models - 100M credits/mo (about 533M tokens)
- 5x the Pro pool
- Research runs and proof vault capacity
Hosted routes use prompt caching and compact run context, so a pool of provider-cost credits goes further than the raw number suggests. One credit is one millionth of a dollar of provider cost. Hosted-model credits buy tokens at the provider's list rate. Pools reset monthly, stop hard at the limit, and never overage. See Privacy.
Arbiter: available on request. Contact kt@ephemerent.com.
Organizations
Governed agent work for engineering teams.
Shared workspaces, pooled credits, Slack requests, approval gates, audit receipts, and centralized policy. An organization pool is shared across seats, reserved before dispatch and settled after, and organization entitlements sit alongside each member's personal plan.
Paid tiers are monthly subscriptions billed in advance via Stripe and renew automatically until canceled. Credits reset each billing month and do not roll over. Hosted model prompts are proxied through Orrery Cloud to provider APIs. Routes, quotas, capacity, and providers may change or be unavailable. Agent output is not guaranteed. See Terms and Privacy.
Questions
Before you download.
What leaves my machine?
When you run a hosted route, the prompt and code context needed for that completion pass through Orrery Cloud to the configured provider. Usage counters store totals and timestamps, and Ephemerent keeps no relay content beyond fulfilling the request. See the privacy policy for the full boundary.
Does Orrery learn from my code?
No. Hosted-model requests are relayed to the provider to fulfill the request and are not captured for training. Local models never leave your machine or network. See the privacy policy for the full boundary.
What can I use for free?
Orrery is free with models on your own hardware or local network, including agent runs, diffs you approve, tests, and receipts. A free account is only needed to download. Pro adds hosted models and credits for $40/month, cancel anytime.
Which platforms are supported?
Windows (x64), Linux (x64 and ARM64) and macOS on Apple silicon (preview). Linux installs update themselves; on Windows and Mac, download the new version from this site. The Mac preview is not notarized yet: open it once, then choose Open Anyway in System Settings, Privacy & Security. The Windows build is not code-signed yet, so Windows may show “Windows protected your PC” the first time: choose More info, then Run anyway.
Can I use Orrery without subscribing?
Yes. With a model server you run yourself (Ollama, LM Studio, vLLM or llama.cpp, on this computer or your local network) the agent is free: pick Your own server in the model menu, no subscription and no API key. Hosted models, deep research on hosted credits, and other providers come with Pro.
Which hosted routes are included?
Pro, Max, and Ultra include monthly credits for hosted models (DeepSeek, GLM and others, served via Doubleword through Orrery Relay) at the amounts shown above. Routes and capacity remain subject to provider availability. Arbiter is not included in plans; it is available on request: contact kt@ephemerent.com.
How is the winning attempt chosen?
By execution, not impression. Each attempt is run against your tests, type-checks, and lint; an automatic reviewer scores them and keeps the winner before merging it back.
Can I cancel anytime?
Yes. Paid tiers are monthly via Stripe and cancel anytime — access runs to the end of the paid period. Credits reset each billing month and don't roll over.
Start free
Open a workspace.
Send one task.
Free with models on your own hardware. Pro adds hosted models for $40/month, cancel anytime, with a receipt for every run.