Penguin is a coding-agent runtime for long-running engineering tasks. It keeps task state, context, checkpoints, tool history, and verification evidence connected across sessions and agents.
uv tool install penguin-aiThen run penguin. First launch creates a workspace and walks through optional model setup.
The objective, context, actions, clarifications, checkpoints, and evidence should remain inspectable after the model stops talking. Penguin makes that runtime state the product.
Sessions persist with checkpoints, rollback, branching, tool history, and file-backed context. Penguin can return to the same engineering state instead of reconstructing it from prose.
Learn moreRun Mode preserves task phases, clarification requests, non-terminal outcomes, and review state so public surfaces do not flatten uncertainty into fake success.
Learn moreSubagents can use isolated or shared context, scoped tool defaults, pause and resume controls, and explicit result reporting through Penguin's message layer.
Learn morePenguin's reliability bar is explicit: implementation, tests, realistic usage, artifacts, and lifecycle state should agree before the work is handed back.
Read about orchestrationChanged files, relevant code paths, and task linkage
Targeted checks first, broader suites when the risk calls for them
Shell, API, browser, or recipe-based exercise of the real behavior
Logs, responses, screenshots, and generated files when they matter
Task phase, clarification state, dependencies, and review truth
The Engine reasons and uses tools; Run Mode owns the task lifecycle; the conversation and event layers preserve what every surface needs to resume, supervise, and review the run.
See the terminal workflowPersist the goal and the task state before the loop runs.
Load project instructions and budget system, reference, dialog, and tool-output context separately.
Read and edit files, run commands and tests, use the browser, or delegate bounded work.
Keep commands, test results, artifacts, failures, and acceptance checks attached to the task.
Finish truthfully: complete, waiting for input, blocked, paused, or ready for review.
Work interactively, automate from a script, connect over HTTP and streaming events, or embed Penguin in Python. Each surface sits over the same runtime services and durable state.
penguinThe full interactive coding workflow: streaming, tools, goals, and session navigation.
penguin-cliScriptable prompts, tasks, configuration, and automation for repeatable workflows.
penguin-webREST, WebSocket, and SSE access for the TUI and your own integrations.
PenguinAgent()Embed Penguin in applications while preserving the same runtime behavior.
First-run onboarding creates Penguin's workspace, verifies it is writable, and can connect OpenAI, Anthropic, OpenRouter, or local Ollama. Model setup can be skipped and resumed later.
Full installation guideuv tool install penguin-aipenguinPython 3.9+ · macOS, Linux, and Windows · local Ollama models supported
Penguin is under active development and available under AGPL-3.0. Start in the terminal, then use the same runtime through the CLI, web API, or Python.