Omnara's January 28th, 2026 post treats its remote-coding app as history and describes the managed-agent platform it now offers; the exact transition date remains unknown.
By [Ryan Merket](https://runtimewire.com/author/ryan-merket)
· Published
Primary source: [Y Combinator](https://www.ycombinator.com/companies/omnara)
Why it matters #
Omnara's transition to managed-agent infrastructure is documented as having happened by January 28th, 2026, though the exact launch date is unknown. Its early usage figures belong to the remote-coding product, leaving adoption of the new control plane the key unanswered measure as it enters a market with named control-plane and execution-layer alternatives.
Ishaan Sehgal and Kartik Sarangmath have taken Omnara from a remote interface for coding agents toward infrastructure for running agents as durable services. A company post dated January 28th, 2026 labels the remote-coding product as historical and says Omnara is now an open-source managed-agent platform. That establishes the transition had happened by that date; the exact launch date is not established by the available evidence.
The Y Combinator profile describes the current product as a model-agnostic control plane with a managed cloud option alongside self-hosting. Omnara's premise is that an agent's identity and history should persist in a control plane while the machines that perform work can come and go. Those environments might be a laptop, VM, container or third-party sandbox.
A founder's infrastructure problem
Ishaan Sehgal says he worked as an AI infrastructure engineer at Microsoft, where he built and led KAITO, a project for running AI workloads on Kubernetes. Before Microsoft, he worked on machine learning at Windsor, a Y Combinator company acquired in 2024, and interned at Roblox. He holds bachelor's and master's degrees in computer science from the University of Illinois Urbana-Champaign, according to his YC profile.
Co-founder and CTO Kartik Sarangmath studied computer science at Georgia Tech, focusing on computer vision research. YC says he published three papers and trained large multimodal models at Meta AI. He later worked at ThirdAI, optimizing neural networks for CPU deployment. Their backgrounds span infrastructure, machine learning and model deployment, relevant experience for a product built to separate the agent's state from its model and execution environment.
Omnara began with a more immediate developer problem: keeping coding agents useful when their users were away from their desks. In the January 2026 post about that earlier product, Sarangmath described developers using agents while walking a dog, commuting or at the gym. Omnara reported that more than 6,000 people had sent over two million messages through the product. Those are company-reported figures for the earlier remote-coding product, not independently audited measures of current usage, paying customers or adoption of the control-plane platform.
The original product let people run coding agents on their own machines and interact with them through web and mobile interfaces. In the current design, the control plane holds agent history and lifecycle information, while connected machines provide the filesystem, operating system, private network or other compute needed for particular tasks. The move expands Omnara's potential customer from an individual developer steering a coding session to a team building agents into its own products and workflows.
State belongs in the control plane
The technical idea is clearest in Sehgal's July 23rd, 2026 post, "Serverless Agents". He argues that an agent should be reconstructible from its durable event history rather than bound to one continuously running process. In the model he describes, a worker takes an event, reads the agent's state, advances the task, records the next action and exits. A later worker can resume when a tool result, approval or new message arrives. The post describes the architecture, but its update saying Omnara has launched does not establish when the launch happened.
Omnara's public repository implements a version of that control-plane thesis. Its documentation says agent state is committed to Postgres and can be recovered after crashes, restarts and temporary machine disconnects. The code is available under Apache 2.0, and teams can run the platform themselves or use Omnara Cloud. Developers can connect compatible model endpoints and select execution options that include their own laptop or VM and supported sandbox providers. The REST API, command-line tools and TypeScript interface let teams build agents into a product, internal Slack workflow or specialized process rather than relying only on Omnara's own interface.
For Sehgal, the argument concerns ownership and reliability. In his post, he writes that the durable log is where an agent's continuity lives, then extends that point to the control plane: a team should be able to own the layer holding its agent history and managing its lifecycle. Omnara's open-source, self-hostable approach gives customers a path to inspect and operate that layer themselves. Its hosted service offers a different tradeoff: less infrastructure for the customer to run, with more reliance on Omnara's cloud. Omnara says the platform itself has no fee: customers can bring their own model keys and machines or self-host for free. The company says hosted models are billed at provider token rates and hosted machines by active time through usage-based credits. That pricing structure leaves costs dependent on the models and execution capacity a customer uses; the published materials do not provide a single all-in price for a production workload. Omnara's pricing information also does not establish comparative operating costs against other platforms.
Self-hosting brings operational responsibility with it. Apache 2.0 code does not itself provide production security, backups, access policies or a reliable deployment; those still depend on how a team configures and operates the system. Omnara's repository explicitly warns that local development defaults are insecure and should not be used in deployed environments.
A crowded layer between agents and compute
Omnara is entering a market where several products describe themselves as control planes or agent infrastructure, though their stated offerings differ. OpenHands Enterprise positions itself as a self-hosted agent control plane, with orchestration, sandboxed execution, observability, policy controls and cost tracking. xpander.ai describes a vendor-neutral control plane for deploying, governing and observing agents. Those are direct points of comparison for teams looking to manage agents across a production environment; the published materials do not establish comparative performance between them and Omnara.
Other providers sit closer to execution than to the control plane. E2B offers isolated microVM sandboxes for agent sessions, while Daytona offers programmable isolated computers and sandboxes. Omnara lists Daytona among the sandbox providers its platform can connect to, placing that service in a complementary role in its documented setup. A control plane that can use outside execution providers has to compete on the orchestration and state layer while still depending on the reliability, availability and cost of the machines underneath it.
That distinction shapes the product's sales challenge. A developer can test a mobile interface on a personal project. A team relying on a control plane for customer-facing agents will also care about failure recovery, permission boundaries, data handling and the work required to maintain the system. Omnara's open-source repository and self-host option address control and inspectability; they do not, on their own, establish enterprise reliability or prove the platform is cheaper to operate than alternatives.
The company's Y Combinator launch listing offered credits to a limited number of teams using hosted models and sandboxes and said Omnara was seeking five design partners for self-hosted, VPC or on-premises deployments. Those terms are recorded in the listing, but they should be read as an early offer, not evidence that the credits remain available or that five deployments were completed.
The earlier traction figures have similar limits. Omnara's six-thousand-user and two-million-message figures show that developers tried its first product; they do not establish current revenue, retention or enterprise demand for the newer platform. YC lists Omnara as an active Summer 2025 company, identifies Y Combinator as its backer, and lists the company in San Francisco with a four-person team. Public materials reviewed here do not establish a valuation, revenue or verified customer count for the managed-agent product.
The bet behind the transition
Omnara's founders have moved from helping a developer steer a local agent to building infrastructure intended to let an agent persist independently of the machine carrying out its work. The first product addressed a human pain point: an agent can stop making progress when it needs input from someone who has left the desk. The current product addresses the systems problem underneath it: preserving the agent's state and deciding where the next action should run.
The design has a boundary. A control plane can preserve instructions, history and pending work while a worker or machine disconnects. It cannot preserve every external condition: a local working tree, a running operating-system process or a credential available only on one device may still require that environment to return or be recreated. Sehgal acknowledges that limitation in his architecture post, which says work dependent on local files may when a device disappears even as the agent remains in the control plane.
Teams will judge Omnara by whether it can resume useful work across interrupted processes and changing environments without turning operations into another project of their own. Sehgal's background building AI infrastructure at Microsoft informs the product thesis. Sarangmath's work across computer vision, multimodal models and CPU deployment reflects a related concern: making the model one component in a larger system. Omnara is trying to provide the layer between an agent's intent and the machines, tools and people needed to carry it out.