India's Soket AI open-sourced LOOP, a Rust harness for AI agents running for weeks

Indian AI research lab and IndiaAI Mission startup building agentic infrastructure.

That in cybersecurity, banking, finance and defence, organisations will choose an open harness they control over closed tools like Codex and Claude Code.

Soket AI, an IndiaAI Mission startup founded by Abhishek Upperwal, built LOOP on the argument that 'a strong model cannot reach peak performance by itself — the longer it runs, the more the harness matters', and chose to open-source the harness rather than keep it proprietary.

The preview, built in Rust for low latency and parallel tool execution, connects to any model via its base URL; developers can branch, pause and resume sessions, preserve history through context compaction, distribute tasks across agents sharing memory, and isolate work from rootless containers up to microVMs. It supports slash-command skills and templates and the Model Context Protocol, including serving developers' own MCP tools.

The agentic coding-and-execution category is dominated by the model owners' closed tools (Codex, Claude Code); an open alternative is a genuine positioning gap.

Regulated workloads — cybersecurity, banking, finance, defence — prize control over infrastructure and data, which open source turns into a product feature.

Model-agnostic by base URL and MCP-compatible (including serving your own tools), it rides whichever models win.

The self-reported efficiency numbers (7.7x less RAM, 17x less local CPU) target a real developer pain: harness resource footprint as agent fleets scale.

When the value moves to the layer above the model, a small lab can compete with the model owners' own tools by open-sourcing that layer and publishing hard efficiency numbers.

As of the 7 October 2026 launch, LOOP was available as a developer preview on GitHub for Linux, macOS and Windows. Soket said it was building sandboxing, personas, multi-agent systems, layered memory, remote agents, trusted remote sandboxes and group chat for agents; the benchmark covered the harness process only and excluded remote inference.