poolsideai/pool: pool is Poolside’s coding agent that runs in your terminal or integrates with any ACP-compatible editor

pool is Poolside’s coding agent that runs in your terminal or integrates with any ACP-compatible editor - poolsideai/pool

Summary

pool is Poolside's coding agent that runs in a terminal TUI, as an ACP server inside compatible editors, or as a non-interactive command for scripts and CI. It supports multiple backends, including Poolside's own service, OpenRouter, Ollama, and any OpenAI-compatible API. The agent handles approvals, file edits, shell commands, and integrates with MCP servers and project-level AGENTS.md instructions.

What it is useful for

This is useful for engineers who want a coding assistant that operates in several modes. In the terminal, it provides an interactive session with slash commands, fuzzy file search via @, shell mode with !, and session resume and rename. As an ACP server (pool acp), it plugs into editors like Zed, JetBrains, and Xcode. As an ACP client (pool --agent-server), it can drive other agents such as the Claude Agent or Codex once the corresponding npm ACP packages are installed. The pool exec subcommand runs a single prompt and exits, which the README explicitly targets at scripts, CI pipelines, and one-off tasks.

How engineers can use it

The documented install is a shell script: curl -fsSL https://downloads.poolside.ai/pool/install.sh | sh on Linux and macOS, with a PowerShell preview for Windows. From a project directory, run pool to start the interactive session, or use pool exec -p "prompt" -o json --unsafe-auto-allow for non-interactive runs. Approvals cycle with Shift+Tab between default, accept-edits, auto, and always-allow modes, and permissions can be declared in YAML with tool and path allow/deny rules.

Hooks run at lifecycle events like PreToolUse and PostToolUse, but the README states they fail open and are not a security boundary, so enforced controls should rely on permissions and sandboxes. Subagents share the workspace, which the README warns can cause conflicts on parallel edits to the same files. Configuration lives in ~/.config/poolside/settings.yaml (global) and .poolside/settings.yaml (per-project), and can define MCP servers, hooks, subagents, and web search providers. For local models, setting POOLSIDE_STANDALONE_BASE_URL and POOLSIDE_API_KEY (for example EMPTY for vLLM) lets pool work against any OpenAI-compatible endpoint; POOLSIDE_STANDALONE_CONTEXT_LENGTH overrides the compaction threshold and must not exceed the server's context window.