Poolside AI

Poolside AI is an AI company focused on building open-weight foundation models and agentic systems for software engineering. Its current platform combines purpose-built coding models, an AI coding agent, developer tools, APIs, and infrastructure designed to let organizations run AI inside their own security boundaries.

Poolside was founded in San Francisco in April 2023 by Eiso Kant and Jason Warner. The company originally focused on building highly capable AI systems for software development, but its current product strategy has expanded toward providing the models and infrastructure needed for enterprise AI agents.

The current Poolside offering centers on the Laguna family of agentic coding models. Poolside’s latest models include Laguna S 2.1 and Laguna XS 2.1. Laguna S 2.1 is designed for long-horizon agentic coding and provides a 1-million-token context window, while Laguna XS 2.1 is a smaller, faster model designed to run on-device.

Poolside models are designed specifically for software engineering rather than general-purpose conversational use. The documentation identifies three major model modes: Agentic, Chat, and Completion. Agentic models can plan and execute multi-step coding tasks, iterate on changes, refactor code, and work with tools. Chat models can answer technical questions and explain code, while completion models provide inline and multi-line coding suggestions.

Agentic coding

Poolside’s coding agent, pool, is designed to operate directly in a developer’s terminal and can also work with ACP-compatible editors. Developers can give the agent a software-engineering task and allow it to inspect files, modify code, execute commands, run tests, and iterate on the result.

This makes Poolside useful for tasks such as debugging, refactoring, implementing features, reviewing code, investigating repositories, and automating repetitive engineering work. Poolside’s documentation specifically describes its models as capable of writing, debugging, refactoring, and automating code.

Developer tools

Poolside can be used through several development environments. Its documentation lists Poolside Assistant, desktop applications, ACP-compatible editors, coding agents, automation platforms, and local model runtimes.

Supported environments and tools include JetBrains, Neovim with CodeCompanion, Zed, VS Code, GitHub Copilot, Cline, Kilo Code, Goose, OpenClaw, Hermes, OpenCode, Pi, GitHub Actions, n8n, Ollama, vLLM Metal, SGLang, and llama.cpp.

Poolside also provides an OpenAI-compatible API, allowing developers to send model requests from their own applications, scripts, and services. This makes the models usable beyond Poolside’s own developer interfaces.

Open-weight models

A major differentiator is Poolside’s focus on open-weight models. The current Laguna family is distributed with model weights under permissive licenses, allowing organizations and developers to evaluate, deploy, and build on the models.

Laguna XS 2.1 is specifically designed to be lightweight enough for on-device operation. Poolside also provides instructions for running models locally using infrastructure such as NVIDIA hardware and local inference runtimes.

The open-weight approach gives organizations more control over deployment compared with AI services that only provide access through a hosted API.

Enterprise and private deployment

Poolside is particularly focused on organizations that require strong control over data, infrastructure, and AI operations. Its platform can be deployed within customer-controlled environments, including cloud VPCs, on-premises infrastructure, and air-gapped environments.

Poolside states that its stack can be deployed inside an organization’s security boundary so that code and data do not have to leave that environment. This positioning is particularly relevant to highly regulated organizations and sensitive government workloads.

For government customers, Poolside highlights air-gapped and disconnected deployments, on-premises infrastructure, GovCloud/edge environments, STIG-hardened environments, and IL5-deployable configurations. It also provides full model weights rather than limiting customers to a hosted API endpoint.

AI agents and automation

Poolside is not limited to interactive coding. Its platform is designed to run AI agents that can operate on software-engineering tasks. The agent runtime can use tools, access files according to configured permissions, operate inside sandboxes, and follow reusable skills.

Poolside supports the Agent Skills format, allowing organizations to create reusable instructions, scripts, references, and templates for specialized workflows. This can help teams standardize how agents perform recurring engineering tasks.

The platform also supports local sandbox environments where organizations can control file-system access, network access, and runtime environments for agents.

Security and governance

Security is an important part of Poolside’s enterprise positioning. The platform provides configurable permissions, sandboxing, secrets management, and controls over tool access.

Poolside’s documentation allows organizations to restrict file access, configure sandbox behavior, manage MCP servers, and define tool permissions. Secrets can also be stored securely and referenced by agents without exposing their underlying values.

This makes Poolside particularly relevant for companies that want AI-assisted development while maintaining control over sensitive source code and infrastructure.

Models and performance

Poolside currently promotes Laguna S 2.1 as its newest frontier-class reasoning model for agentic coding. It contains 118 billion total parameters, 8 billion active parameters, and a 1-million-token context window.

Laguna XS 2.1 contains 33 billion total parameters and 3 billion active parameters, with a 256K context window. Its smaller architecture makes it suitable for faster workloads and on-device deployment.

Poolside’s models are trained specifically for software-engineering workflows and are intended to perform multi-step tasks rather than simply generate isolated code snippets.

Pricing

Poolside does not currently publish a conventional public subscription pricing table for its enterprise platform. Enterprise deployments are handled through direct engagement with Poolside and can involve customer-controlled cloud, on-premises, or air-gapped infrastructure.

Poolside’s government offering emphasizes that agencies running models on their own hardware do not pay per-token inference fees; instead, the economics are based around the deployment and infrastructure model.

For developers wanting to experiment with Poolside models, the current website provides free access options. Poolside’s model page states that Laguna models are free to use for a limited time, while the documentation provides a free developer-access path for the Poolside Agent CLI.

Poolside models are also available through third-party platforms such as OpenRouter and Vercel AI Gateway, where pricing is determined by those providers rather than representing Poolside’s enterprise subscription pricing.

Review

Poolside AI is a compelling option for organizations looking for enterprise-grade agentic coding combined with greater control over AI infrastructure. Its strongest differentiator is the combination of open-weight coding models, agentic development capabilities, local deployment, and enterprise infrastructure options.

The Laguna models are purpose-built for software engineering and support long-context, multi-step coding workflows. Poolside also gives developers considerable flexibility in how they access the models, including its own coding agent, desktop and editor integrations, APIs, OpenRouter, Vercel AI Gateway, and local inference environments.

Another major advantage is its emphasis on security boundaries. Organizations in financial services, defense, government, and industrial environments can deploy the technology within their own infrastructure rather than sending sensitive source code and data to a conventional external AI service.

The main limitation is that Poolside is primarily targeted at professional software engineering and enterprise deployments, rather than being a general-purpose AI assistant. Enterprise pricing is also not publicly standardized, making direct cost comparisons with per-seat coding assistants difficult.

For engineering organizations that prioritize control, privacy, customizable deployment, and agentic coding, Poolside is a strong alternative to conventional cloud-only AI coding platforms.