PI

At a Glance:

Pi Agent Harness is a monorepo providing an interactive open-source CLI coding agent, a tool-calling agent runtime with state management, and a unified multi-provider LLM API, designed for local use with explicit containerization guidance for sandboxing.

Overview:

Pi Agent Harness is an open-source monorepo that contains a self-extensible, interactive coding agent CLI, an agent runtime for tool calling and state management, and a unified API for multiple LLM providers. The project is built for developers who want to use a local, terminal-based coding agent. It explicitly states that it does not include a built-in permission system, instead offering documented patterns for containerization and sandboxing to enforce filesystem, process, or network boundaries. The monorepo also provides a terminal UI library and links to a separate package for Slack and chat automation workflows.

Key Decision Points:

• No built-in permission system: Pi runs with the user's process permissions by default, so users requiring filesystem, network, or credential restrictions must implement separate sandboxing.

• Sandboxing is documented, not bundled: The project provides explicit patterns for using OpenShell, a local Linux micro-VM via the Gondolin extension, or plain Docker to isolate the agent, rather than integrating a permission layer.

• Supply-chain hardening is a core practice: The repository treats dependency updates as reviewed code, using exact version pinning, a package-lock.json ground truth, a pre-commit block, and a shrinkwrap generation allowlist for lifecycle scripts to manage risk.

• Terminal-native interaction: The primary interface for the coding agent is a CLI, built with a custom terminal UI library for differential rendering.

Core Features:

• Interactive coding agent CLI: An open-source tool (@earendil-works/pi-coding-agent) for interactive code generation and task execution from the terminal.

• Agent runtime with tool calling: A core runtime (@earendil-works/pi-agent-core) that manages agent state and the execution of tool calls.

• Unified multi-provider LLM API: A library (@earendil-works/pi-ai) that provides a single API to interact with models from OpenAI, Anthropic, Google, and other providers.

• Terminal UI library: A dedicated library (@earendil-works/pi-tui) for building terminal user interfaces with differential rendering.

• Session sharing for OSS work: Users can record and publish their coding agent sessions using an external tool (badlogic/pi-share-hf) to contribute real-world task data for improving coding agents.

• Supply-chain verification scripts: An npm run check script verifies pinned direct dependencies, TypeScript import compatibility, and the integrity of the coding-agent shrinkwrap file.

Use Cases:

• Developers wanting a local CLI coding agent: For users who prefer an interactive coding agent that runs directly in their terminal using their own LLM provider API keys.

• Security-conscious users implementing custom sandboxing: For developers who need stronger execution boundaries and are prepared to choose and set up a containerization pattern as documented by the project.

• Open-source contributors sharing agent telemetry: For users who want to contribute real coding session data to the community to help improve agent models and tools beyond synthetic benchmarks.

Open-Source Alternative Value:

As an open-source project, Pi Agent Harness provides transparent access to the source code of its agent runtime, CLI, and LLM abstraction layers, allowing developers to inspect, modify, or extend the tool's internal tool calling and state management logic. Its explicit separation of the agent from a built-in permission system, paired with documented sandboxing patterns, gives users a clear architectural choice in security boundaries. The project's methodical approach to supply-chain hardening—treating npm dependency changes as reviewable code—offers a concrete and transparent dependency management model for a coding agent.