About
We built the AI, the editor, and the infrastructure underneath.
Most AI coding tools are an interface on top of somebody else's model. MonkeysCode is the model, the agent runtime, the serving stack and the editor, built by the same team. That is a harder way to do it. It is also the only way to give developers a tool that does not answer to another company's roadmap or pricing.
One product, three surfaces, one engine.
MonkeysCode is an agentic development environment. You describe what you want, and an agent plans it, edits across files, runs your commands, executes your tests and corrects its own mistakes before handing the work back.
The IDE
A full editor with the agent integrated rather than bolted on. Where most people start.
The Terminal
The same engine, headless. Interactive in your shell, non-interactive in CI, and able to keep working on a schedule while you do something else.
Troop
Orchestrates many agents at once across separate projects, each in an isolated workspace, each producing a signed diff and a test report you review before anything merges.
They share one runtime. The agent behaves the same way in all three, because it is the same agent.
Our own model, not a reseller agreement.
Capuchin is the AI at the centre of MonkeysCode. We train it, we serve it, and we run the GPUs it lives on.
It is specialised for software engineering rather than general conversation, which does two things at once. It performs better on real coding work than a general model of a comparable serving cost. And it costs a fraction as much to run, which is what lets us include generous usage in a flat subscription instead of metering every token.
It runs in two modes the editor switches between automatically. Flash answers in under a second and handles the large majority of everyday work: completions, applying edits, quick questions. Reason thinks step by step, within a budget, for the harder minority: planning changes across many files, debugging, weighing tradeoffs.
92.7%
of requests served by Capuchin
In production, the frontier models handle the difficult remainder — which is exactly what they should be doing.
Three decisions that shaped everything else.
Decision 01
Your code stays on your machine by default.
The codebase index is built and searched locally. Nothing is uploaded unless you turn that on. Air-gapped deployment, with a local index and a local model, is a supported configuration rather than a theoretical one. That matters because there is an entire category of developers, in finance, healthcare, defence and government, that no major AI coding tool can legally serve.
Decision 02
You choose the model, always.
Capuchin by default. Claude, Gemini or ChatGPT when the task calls for it. Or a model running on your own hardware. Switching is a setting, not a migration. We built our own model specifically so we could afford to support the ones we did not build.
Decision 03
Every agent action is verifiable.
Runs are signed and replayable. A change that fails your test suite is discarded rather than applied. You can hand a run to a teammate, or an auditor, and they can confirm exactly what happened without a live session and without taking anyone's word for it.
A bill you can forecast.
Flat subscription. No per-token surprises.
When this category moved to usage-based billing, developers reported costs rising by an order of magnitude on agentic workloads. The problem was never the absolute number. It was that nobody could predict it.
Quota is enforced per model on a rolling five-hour window, sized to your plan. When a window is exhausted, nothing bills you by surprise. Work continues on the other pool, or resumes when the window rolls. There is a hard spend cap, and bring-your-own-key traffic is paid directly to your provider with no markup from us.
We would rather charge you for the tool than take a cut of your model bill.
Jorge Peraza, Founder.
Jorge Peraza
Founder · MonkeysCloud
Jorge Peraza is the founder of MonkeysCloud and the creator of MonkeysCode and Capuchin.
He has spent more than twenty years building production systems, and the last several building the infrastructure that AI development tools run on. That combination is unusual and it is the reason MonkeysCode exists in the form it does: the same person who designed the agent runtime also sized the GPU capacity it runs on and wrote the admission control that keeps it economical.
Jorge began his career in Costa Rica, studying Computer Software Engineering at ULACIT and working as a senior engineer at Gorilla Logic in Escazú. He went through the AUGE incubator at the University of Costa Rica and organised DrupalCamp Costa Rica in 2014, one of the country's first gatherings of that community. He later moved to the United States, where he worked as a senior full-stack and platform engineer at Tesla, building and supporting high-traffic Drupal and React systems for enterprise web platforms, and at Cidi Labs as a senior Symfony developer. Across two decades of consulting and product work before that, he delivered web platforms, APIs, CMS implementations and backend systems for agencies, enterprises and product teams across a range of industries.
He founded MonkeysCloud in 2019 and spent the following years building a vertically integrated developer platform: repository management, deployment workflows, server provisioning, CI/CD integration, a PHP framework, a content management system and a transactional email API. That work established the pattern MonkeysCode follows, and more importantly it established MonkeysAI, a self-hosted inference layer with fast and reasoning execution paths built on open-weight models. Building the serving platform first is what made building a model realistic.
MonkeysCode and Capuchin followed. Jorge built the product from zero: the TypeScript agent runtime, the mcode command-line interface, the multi-agent desktop orchestrator, a Go proxy handling quota and admission control with burst and weekly governors, and the GPU serving stack on committed H100 and H200 infrastructure running SGLang. He modelled the capacity himself, at roughly seven thousand concurrent users per node, and designed the dual-pool quota architecture and nine-tier packaging that make a flat subscription economically viable against frontier model costs. The reinforcement learning pipeline that specialises Capuchin further, using GRPO and DPO on real usage trajectories, is his work as well.
His technical range spans Python, TypeScript, Go and PHP, with hands-on depth in model serving through vLLM and SGLang, self-hosted open-weight models including Llama, DeepSeek and large mixture-of-experts checkpoints, embeddings and retrieval, model routing, GPU scheduling under concurrency pressure, and LoRA and QLoRA fine-tuning. On the infrastructure side, Terraform, Kubernetes, GKE, Docker and the reliability discipline that goes with running inference endpoints in production: latency percentiles, structured logging, startup probes, capacity planning and cost attribution.
He is based in Denver, Colorado.
MonkeysCode is the fifth product, not the first.
Each works on its own. Together they cover the path from writing code to running it in production, built on one set of conventions by one team.
MonkeysCloud
A developer platform covering repositories, deployment workflows, environment configuration and server orchestration.
MonkeysLegion
A PHP 8.4 framework with AI orchestration built into it.
MonkeysCMS
Content management for content-heavy work.
MonkeysMail
Transactional email API for application-generated messaging.
MonkeysAI
The self-hosted inference layer that became the foundation Capuchin was built on.
That history is the reason MonkeysCode exists in the shape it does. Building the inference platform first is what made building the model plausible.
Denver, Colorado.
MonkeysCloud is based in Denver, with engineering distributed across the United States and Latin America. We are a small team by design. Owning the model and the infrastructure means the cost structure works at small scale, which means we do not have to grow before the product is right.
Try it, then tell us what broke.
The IDE is available for Windows, macOS and Linux. The Terminal and Troop follow shortly. Thirty days of Capuchin is free, with no credit card.
We are pre-1.0 and we say so. If something is wrong, the founder reads the replies.