Spacebar unveils shared runtime for real-time multimodal AI agents
Ayush Agarwal's team is offering developers a persistent canvas, shared browser and permission layer for agents that work alongside people in real time.
By Ryan Merket · Published
Why it matters
A public endorsement from veteran engineer and investor Gokul Rajaram highlights Spacebar.ai’s early traction in the crowded AI productivity market, potentially unlocking investor interest and signaling a credible product milestone.
Ayush Agarwal and co-founder Amogh Asgekar have unveiled Spacebar, an infrastructure platform designed to let humans and AI agents work together on a persistent, synchronized canvas. The company is positioning the system as a model-agnostic runtime for real-time multimodal products, rather than another standalone AI assistant.
Spacebar combines audio, video, cursor activity, embedded applications, browser sessions and canvas objects in one event stream. Human participants and agents operate under the same permission model, while developers can connect in-house models and keep model weights on their own infrastructure.
The founders are targeting a gap between increasingly capable multimodal models and the software surfaces available to them. Most collaboration products place AI beside the work as a chatbot, not inside the same stateful environment. Spacebar instead gives agents a persistent place to observe events, act on objects and return to unfinished work across sessions.
A shared browser and canvas for agents
One of the platform's core components is a shared browser that a human and an agent can operate simultaneously. Spacebar says the agent receives continuous video, live DOM access and interaction events rather than relying only on periodic screenshots or accessibility trees.
Chrome and Firefox extensions bring that agent layer into a user's existing browser profile, preserving cookies, authenticated sessions and two-factor authentication. The approach is intended to avoid the remote virtual machines commonly used by computer-control agents, which can force users to recreate their working environment.
The broader canvas supports documents, whiteboards, video tiles, server-hosted applications and virtual machines. Objects retain their positions between sessions, and every state change can be observed or addressed through the same SDK. Agents inherit the roles and access rights of the users they represent, according to the company, rather than receiving a separate set of elevated controls.
Two agent deployment models
Spacebar is offering two primary patterns for developers. Real-time agents can join live sessions over WebRTC, listen to per-participant audio, watch changes on the canvas and speak or act while the session continues. Always-on agents can work between sessions or respond to scheduled and event-based triggers, leaving their sources, drafts and other outputs arranged in a persistent space for a human to review.
The company has also built a canvas-native assistant with 24 structured tools for creating boards, organizing folders, controlling devices, inserting applications and adding generated material. Tool calls pass through the same authorization layer used for human actions.
For developers, the platform exposes an SDK, REST API, webhooks and an MCP server. Its speech-to-text, language-model and text-to-speech providers are configurable by session, allowing builders to use different vendors or their own models.
Performance and production claims
Spacebar says its event propagation is 50 milliseconds at p99 and that availability has remained above 99.99%. It attributes that performance to a custom CRDT engine, dedicated worker pools, layered caching, asynchronous snapshot compaction and per-space ownership locks.
Those figures are company-reported and have not been independently verified. Spacebar says the underlying system has powered paid workloads for Pencil Spaces for more than four years and has served more than 200 million minutes. It also claims SOC 2, HIPAA and GDPR compliance, with regional infrastructure and JWT-scoped real-time channels.
Rajat Goel, a senior engineer at Thinking Machines Lab, described Spacebar as the substrate on which the next generation of AI products will be built, citing its multi-user, multi-agent architecture and production scale.
The unveiling also drew attention from Gokul Rajaram, who responded to Agarwal's post on July 21: "This was eye opening (and mind blowing)! Congrats Ayush and Spacebar team." The response received 47 likes, four reposts and more than 21,000 views within hours, according to the public post metrics cited in the original report.
The business case for builders
Agarwal's team argues that building an equivalent multiplayer canvas, full-duplex media layer, shared browser, agent authorization system and persistent memory stack would require 18 to 24 months. That estimate is part of Spacebar's sales case rather than an independently tested benchmark, but it clarifies the target customer: AI companies that want to build an application on top of collaborative infrastructure instead of maintaining the underlying real-time systems themselves.
Spacebar plans to monetize through usage-based infrastructure pricing, enterprise licenses and trajectory-data licensing for frontier AI labs. It is also offering integration, data and white-label product partnerships. Public pricing and a detailed customer list were not disclosed.
The platform will not be necessary for every AI application. Single-turn voice tools, basic chatbots and headless batch agents may need little more than an inference endpoint. Spacebar is making a narrower bet that products involving multiple participants, long-running tasks, live browsers or visually structured work will require a persistent surface where people and models can see and change the same state.