Plandex is an open-source AI Coding Agent that runs in the terminal. The project describes it as a development tool that can plan and execute large coding tasks spanning many steps and touching dozens of files, and states that it is designed for large projects and real-world tasks. Its README positions it against tools it says struggle once a project grows past a certain size or the changes become too complex.
It is distributed as a one-line, zero-dependency CLI install together with a server that can be run
locally in Docker or hosted on the user’s own machine. The CLI is invoked as plandex, with pdx
as a short form, and the repository is published under the MIT license. Windows is supported only
through WSL — the project states Plandex does not work correctly in the Windows command prompt or
PowerShell.
Capabilities
Context handling is the capability the project puts first. It states an effective context window of 2M tokens with the default model pack — roughly 100k per file — and says Plandex loads only what each step needs. Directories of 20M tokens or more are indexed using tree-sitter project maps, which the project also uses for syntax validation and says covers more than 30 languages. Context caching is used for OpenAI, Anthropic and Google models, which the project says reduces cost and latency.
Changes are staged rather than written straight into the project. The README calls this a cumulative diff review sandbox: AI-generated changes are kept separate from the project’s files until they are ready to apply, and command execution is controlled so that a run can be rolled back and debugged. The project presents this as the mechanism that keeps an agent from leaving a mess behind in a codebase.
Autonomy is configurable rather than fixed. The project states Plandex is capable of full autonomy — loading relevant files, planning and implementing changes, executing commands and debugging automatically — while also offering fine-grained control and a step-by-step review process where that is wanted. Automated debugging covers terminal commands such as builds, linters, tests, deployments and scripts, and, where Chrome is installed, browser applications as well.
Alongside implementation there is a project-aware chat mode, which the project describes as a way to flesh out ideas before moving to implementation and to ask questions about a codebase. Models from several providers can be combined, with curated model packs offered as different tradeoffs of capability, cost and speed, including open-source and provider-specific packs. The project describes its file edits as prioritising correctness, validating both syntax and logic as needed with multiple fallback layers when problems arise.
Every update to a plan is placed under version control of its own, including branches for exploring multiple paths or comparing different models. This sits alongside git integration with commit message generation and optional automatic commits. The developer-facing surface is a REPL with fuzzy auto-complete for commands and file loading, plus a CLI for scripting or piping data into context.
Adoption & Ecosystem
Plandex is run against a model provider the user supplies. The project documents an OpenRouter.ai key as one option and points at other model provider accounts and API keys as alternatives, and states that a Claude Pro or Max subscription can be connected for calls to Anthropic models, with the prompt to connect one appearing on first run.
Of the two hosting options the README lists, the managed one is being retired: it states that Plandex Cloud is winding down as of 10/3/2025 and is no longer accepting new users. Self-hosted and local mode — Docker locally, or the server on the user’s own infrastructure — is the remaining route.
The project is developed in the open on GitHub, with a Discord server, GitHub Discussions and an issue tracker named as the places for feedback and questions, and documentation hosted separately at docs.plandex.ai.