Cosmic Stack sits squarely in the agent stack. Mercury Agent is a coding agent with MCP support, a provider-agnostic runtime, and a persistent memory architecture the company treats as its core technical bet. Lazy Frames is explicitly built to be driven by other agents through a skill.md interface, so any coding agent can pick it up without a bespoke integration. Cosmic Stack Cloud is a managed runtime for long-horizon agent workloads, with checkpointing, tracing, and bring-your-own model providers.

The company's public research pushes a specific argument into the ecosystem: that memory architecture, not model quality, is what makes an agent product durable, and that current agent benchmarks mislead. For people building or operating agents, Cosmic Stack matters as an open-source alternative to closed coding agents and as a vendor arguing that the runtime layer should be open and self-hostable. The connection is direct and central, not tangential.

A lab with a product surface

Cosmic Stack is a research and product lab, founded in 2025 and remote-first, that builds AI agents, the infrastructure to run them, and the developer tools around them. The pitch on its homepage is blunt: frontier inference got two orders of magnitude cheaper, but most "AI products" are still thin chat wrappers. Cosmic Stack's bet is that the gap between what's technically possible and what's actually shipped is the opportunity, and that the agent runtime, not the model, is where defensibility lives.

Four surfaces, one pipeline

The company organizes itself into Labs, Open Source, Incubator, and Cloud, and describes them as stages of one pipeline rather than separate business lines. Labs prototypes against real evals and most ideas die there in public. Open Source hardens the survivors. The Incubator turns open-source primitives into founder-led companies. Cloud runs the whole stack as a managed service. The company calls this pipeline its moat, with each stage feeding the next production evidence and the previous one problems worth researching.

Mercury Agent

The flagship is Mercury Agent, a terminal-native coding agent for professional engineers. It ships 40+ tools, an Autopilot mode that plans and recovers across hundreds of turns unattended, a persistent "Second Brain" memory, and a typed Skills library with 128 community and 23 built-in skills. The runtime is provider-agnostic across OpenAI, Anthropic, Cohere, Mistral, xAI, and OpenRouter, supports MCP, and installs with a single npm command. It is MIT-licensed, self-hostable, and runs with no telemetry or account. The GitHub repo sits around 3,500 stars, and the broader org has roughly 4,200 across its projects.

The research behind the product

Cosmic Stack publishes research that reads like engineering postmortems. One post, drawn from running its coding agent through 1,000+ turn tasks, argues the real failure mode in long-horizon tool use isn't reasoning but memory shape, and that hierarchical scratchpads and failure-weighted retention matter more than model quality. Another takes aim at agent benchmarks, arguing leaderboards measure task completion under generous constraints that don't survive contact with production. The company's stated principles follow from this: evals before features, open source by default, and small teams with end-to-end ownership.

Beyond the coding agent

Two more products ship from the same lab. Sav.ink is a local-first personal AI finance manager in private beta for web and iOS, aimed at people with a dozen accounts and no patience to reason about them; pricing is targeted at a single $20-40 monthly subscription. Lazy Frames is an agentic video tool that takes a skill URL, has an agent author and render a spec, and produces byte-identical MP4s fully locally. Both reflect the company's local-first stance on sensitive data.

The managed runtime

Cosmic Stack Cloud is the commercial layer, in early access with a Q3 2026 GA target. It runs the same runtime as open-source Mercury, adds checkpointed resumable state for multi-hour jobs, built-in tracing, bring-your-own model providers, and SOC 2 / HIPAA with US, EU, and APAC data residency. Pricing is per successful task rather than per CPU-minute. The competitive question is whether a small lab can win developer mindshare against better-funded coding-agent incumbents; the open-source distribution and the pipeline structure are its answer.