Remoroo
Remoroo is an autonomous experimentation engine for robots and code: give it a goal and a metric, and it practices until the target is met, then proves the result on held-out tests with a measured, auditable success rate.
What is Remoroo?
Remoroo is an autonomous experimentation engine sold in two shapes that share a single agent loop.
On the robotics side, you bolt down a rig and wire its arms and cameras to a machine, then run four commands. Remoroo discovers the hardware, builds its connection layer, models the cell, calibrates the cameras against the arms, generates collision geometry and commissions whole-robot motion, showing each gate live in 3D for you to approve before the next one starts. You then describe the job in a single sentence, such as clear the table into the black box, and that sentence is the whole interface: no teach pendant, no scripting, no cell programming and no teleoperation. The robot practices on the real cell, attempt after attempt, until the task works. A skill only counts as done once it passes held-out verification, meaning fresh start states it never practiced, scored by a pinned judge, with a measured success rate you can audit. When the cell changes and that rate drifts, the skill re-trains and re-verifies itself, and re-verification is always free.
On the software side, the same loop drives a command-line engine for ML and deep-tech repositories. You supply a goal and a metric; the agent captures a baseline, plans, edits, re-runs the evaluation through a metric_gate tool and iterates until every constraint holds. Execution stays on your machine in a Docker or virtualenv sandbox, while a hosted Brain does the reasoning and a Control Plane routes jobs through queues without ever touching your code. A demand-paging memory system borrowed from operating-system virtual memory keeps the agent coherent across more than two hundred tool calls: results are pinned, paged out to compact retrieval handles or dropped, a persistent key-value note tool survives every compaction, and cross-run recall carries lessons from one run to the next. A wall-clock budget in hours ends the run with a verdict instead of a kill.
Every run leaves artifacts inside your repository: a report, a clean patch, a replayable trace, checkpoints and metric files.
What it does
- Commission a robot cell on its own: find the arms and cameras, calibrate, model collision geometry and bring up safe whole-robot motion
- Train a robot on a task described in one sentence, by practicing on the real hardware with nobody driving
- Prove a skill on held-out tests it never practiced and report a measured, auditable success rate
- Run a goal-and-metric experiment on a code repository until every constraint passes, then hand back a clean patch
- Refactor across hundreds of files, migrate dependencies or standardize patterns under a check that must pass
- Optimize runtime, memory, container image size or training hyperparameters against baseline-relative targets
- Detect drift when a cell changes, then re-train and re-verify the skill at no extra cost
When to use Remoroo / When not to
A quick filter to help you decide if Remoroo is the right fit.
When to use Remoroo
- ML and deep-tech teams that already score their work against a fixed metric harness
- Manufacturing and automation engineers with dexterous cell tasks still done by hand
- Platform, DevOps and data engineers facing large refactors, dependency migrations or hard performance targets
- Research groups running long overnight experiment sweeps on hardware they control
- Engineering leaders who want an auditable success rate rather than a vendor's claim
When not to use Remoroo
- Teams whose goals cannot be reduced to a measurable pass or fail
- Anyone shopping for a conversational coding assistant or autocomplete: Remoroo states plainly that it is not a coding agent
- Users without a terminal, Python 3.10 or later, and Docker or a virtualenv sandbox
- Organizations that need a documented GDPR posture, a Data Processing Agreement, a subprocessor list or a stated hosting location
- Small budgets: past the free first robot and first skill, a verified skill is priced at 6,000 US dollars
How to use Remoroo
A typical end-to-end flow, from setup to results.
- Install the command-line tool with pip install remoroo on Python 3.10 or later
- Authenticate with remoroo login, which opens a browser once and stores credentials locally
- Make sure Docker is running, or plan to pass --engine venv to use a virtualenv sandbox instead
- For software work, launch remoroo run --local with a --goal string, a --metrics string and a wall-clock --budget in hours
- Follow the run in the full-screen terminal interface: p pauses, Ctrl+D detaches while the run continues, q quits and aborts
- Collect the result under .remoroo/runs/ in your repository, apply it with git apply on the final patch, then review with git diff
- For robotics, wire the arms and cameras to the machine, then run remoroo robot activate once for that rig
- Run remoroo setup to open the Studio in your browser and approve each commissioning gate in 3D
- Teach a job with remoroo task followed by one sentence describing what the robot should do
- Read the held-out verification result and leave the skill under monitoring so drift triggers a free re-verification
Pros & Cons
Pros
- The deliverable is a proof rather than a promise: a measured success rate on tests the robot never practiced
- Code stays on your machine, since execution is local by default and the Control Plane is described as never touching the repository
- The command-line tool and offline engine are open source and publicly auditable on GitHub
- The first robot and the first skill cost nothing, and re-verification of a skill you own remains free
- A 90-day guarantee backs every paid verified skill
- The public architecture write-up is unusually detailed for a product at this stage
- Full run artifacts, including a report, a clean patch, a replayable trace and checkpoints, make the work reviewable after the fact
Cons
- No mention of the GDPR anywhere, no Data Processing Agreement, no subprocessor list and no stated hosting country
- The privacy policy collects run data to refine Remoroo's own planning models, with no documented way to opt out
- No legal entity with a corporate form and no postal address are published
- Agent credit plans are described on the pricing page but cannot be bought: none are available
- Hosted execution does not exist in the current release, and --remote exits with a sales message
- Two documentation sections, Engine Architecture and Security and Guardrails, are still empty
- Very young: the domain was registered in December 2025 and the job listing describes a pre-seed team
Pricing & Plans
A permanent free entry point is offered: the first robot is commissioned at no charge and the first skill is trained and verified free of charge, with re-verification of any skill you own remaining free thereafter. Beyond that, the lowest published paid price point is USD 2,500 per robot per year, applying from the second robot onward and covering setup, monitoring, drift recovery and free re-verification. Each additional verified skill is priced at USD 6,000 and carries a 90-day guarantee. Fleets of twenty robots or more move to a flat fleet license whose amount is not published. All figures are quoted in US dollars.
- connection
- calibration
- safety and motion
- commissioned by the agent in the Studio
- the robot trains itself on your first task and passes the same verification as any paid skill
- training until the task works
- then held-out proof on your own cell with a measured success rate
- backed by a 90-day guarantee
- a full year of setup
- monitoring
- drift recovery and free re-verification of every skill that robot runs
- flat fleet license
- amount not published
- Starter (150 credits) and Pro (330 credits)
- prices not published and currently unavailable for purchase
Data, GDPR & hosting
A consolidated view of how Remoroo handles your data.
GDPR overview
There is no GDPR implementation to report. The word GDPR appears nowhere on the site, in neither the terms nor the privacy policy, and a full-text search across every page collected, raw HTML included, returned nothing on the GDPR, the CCPA or data protection law generally. No Article 27 EU representative is designated, no Data Processing Agreement is published or offered on request, no subprocessor list exists, no hosting country or region is named, and no retention period is stated. The only privacy contact published is privacy@remoroo.com. The terms are governed by the law of the State of Delaware, USA, with exclusive jurisdiction in Delaware courts. The privacy policy runs to six short sections and reads as a product commitment about local execution and code ownership rather than as a compliance notice, so a European buyer has nothing to rely on here.
Who owns the data?
Both documents are dated 5 April 2026 and put ownership squarely with the customer. The privacy policy states that you own 100% of the code you develop, optimize or repair with Remoroo, and that Remoroo claims no intellectual property over your project files, ideas or patches. The terms say the same: you retain your code, data and materials, and grant Remoroo only a limited license to host, process and display what you submit, as far as needed to provide and improve the service. Remoroo keeps the rights to its own product, branding and documentation. Because execution is local by default, raw source code stays inside your infrastructure unless you opt in to the hosted Cloud Engine.
Reuse rights
Whatever the engine produces is yours to reuse without asking: patches, reports, traces and metric files belong to you, and Remoroo claims no rights over them. The limited license granted in the terms runs the other way, covering only the hosting, processing and display of what you submit, as far as needed to run and improve the service. On its own side, the privacy policy says Remoroo collects selective data points from local runs, namely metrics, targeted source files described as hydrated context, final patches described as automatically sanitized, and non-sensitive execution logs, used exclusively to refine its autonomous planning models. No opt-out from that collection is documented anywhere on the site.
Data retention & training
Hosting summary
No hosting jurisdiction is declared. The privacy policy says that by default all code execution happens locally, on your own machine or in a dedicated container, and that raw source code never leaves your infrastructure unless you explicitly opt in to a hosted Cloud Engine whose location is not stated. The architecture write-up splits the system in three: a terminal interface and a worker on your machine, a Control Plane that routes execution requests through queues and is described as never touching your code, and a remote Brain that runs the reasoning loop. Authentication uses a session token or an API key. Run artifacts stay in your own repository, with a possible cache copy in your home directory, and credentials sit in a local configuration file. The website itself resolves to an anycast address operated by a United States content delivery network and uses two third-party analytics services. Beyond that, no country, no region, no data center and no subprocessor list are published, so where the hosted components actually run remains unknown.
Things to keep in mind
Risks and trade-offs to weigh before adopting Remoroo.
- Run data leaves your side: the privacy policy collects targeted source files, sanitized patches and execution logs to refine Remoroo's models, and no opt-out is documented
- A measured success rate on held-out tests is a strong signal, not a safety certificate; a robot that passes verification can still fail on a state nobody thought to test
- The metric quietly becomes the goal, and an agent optimizing hard against a badly written metric will satisfy it in ways nobody intended while the patch still looks clean
- Delegating overnight experiment sweeps is habit-forming: a team can keep the verdict and slowly lose the intuition about why a change actually worked
- Liability is capped at the greater of twelve months of fees or one hundred US dollars, and fees are non-refundable unless stated otherwise
- Activating a rig is permanent, since identity is derived from the wired hardware, so a wiped disk or a replaced camera never re-registers the same robot
- With no published legal entity, address or data-protection position, there is little to fall back on if a dispute or a privacy question arises
Setup & Integrations
Technical difficulty
Developer-level, and terminal-only on the software side. You need Python 3.10 or later, Docker or a virtualenv sandbox and preferably a Git repository; installation is a single pip command and signing in opens a browser once. The real difficulty is not installation but writing a metric the agent cannot game: the documentation devotes a section to baseline-relative and multi-metric targets and lists five common failure modes with fixes. On the robotics side Remoroo does the commissioning itself, so you wire the arms and cameras and approve each setup gate in 3D. No cell programming, scripting or teach pendant is required.
Deployment
Integrations
Behind Remoroo
Fundraising
Resources
All the official URLs gathered for verification and reference.
Frequently asked questions
What exactly am I buying?
What does a verified skill cost?
Does my source code leave my machine?
Who owns the code Remoroo produces?
Does Remoroo use my data to train its models?
What do I need to run it?
Can Remoroo run the workload on its own infrastructure?
What happens when the cell changes and a skill degrades?
Is Remoroo GDPR compliant?
Is there a mobile app?
Should you pick Remoroo?
Remoroo sells something unusual for an AI product: a proof rather than a promise. Whether the job is a robot arm clearing a table into a box or a data pipeline that must run in under two seconds, the run only ends when the result passes a check the system did not choose for itself, whether that is a set of held-out start states the robot never practiced or an evaluation harness the agent is forbidden to modify. That discipline runs through the whole design, and the public architecture write-up is candid about how the agent stays coherent across hundreds of tool calls, where the code actually executes and what the hosted side is allowed to see.
The engineering reads as credible and the commercial terms are legible. Local execution by default, an open-source command-line tool, a free first robot and first skill, free re-verification for life and a 90-day guarantee on every paid skill all point the same way.
The gap is everything around the product. There is no legal entity with a corporate form, no postal address, not one mention of the GDPR, no Data Processing Agreement, no subprocessor list, no stated hosting country and no retention period, while the privacy policy openly collects targeted source files, sanitized patches and run logs to refine Remoroo's own models with no documented way to opt out. The domain was registered in December 2025 and the company describes itself as pre-seed.
Read it as an early-stage tool for teams that can already express success as a number and are comfortable running everything on their own hardware. For those teams the verification model is genuinely differentiating. For anyone with a procurement or data-protection review to clear, the paperwork does not yet exist.
- Choosing a selection results in a full page refresh.
- Opens in a new window.