Plandex
Type: full-code · Vendor: Plandex AI · Language: Go · License: AGPL-3.0 / MIT (mixed; see repo) · Status: active · Status in practice: emerging
Open-source terminal AI coding agent for large projects: stages edits and command execution in a cumulative diff-review sandbox, supports configurable autonomy from full-auto to step-by-step, and version-controls every plan with branches.
Description. Plandex is a Go terminal agent aimed at multi-file, large-scope coding tasks. Its central idea is a sandbox of tentative changes: AI-generated diffs sit outside the project files until reviewed and applied. Plans are first-class objects with their own version history and branches, so multiple paths or model comparisons can run side by side. Autonomy is a slider - None, Basic, Plus, Semi-Auto, Full-Auto - controlling what the agent may do without confirmation. A REPL with fuzzy completion and tree-sitter-driven project maps round out the loop.
Agent loop shape. Plan-and-execute over a tentative-diff sandbox. The user gives a task; Plandex builds a project map, drafts a plan, generates diffs and commands inside the sandbox, optionally runs commands with rollback, and only applies edits to project files when the user (or the configured autonomy level) accepts them. Plans are versioned; branches let multiple plans coexist.
Primary use cases
- long-running multi-file refactors in a terminal
- tentative-diff review before files mutate
- configurable autonomy from step-by-step to full auto
- branching plans for exploring alternatives
Key concepts
- Cumulative diff-review sandbox → approval-queue (docs) — AI-generated changes stay separate from project files until applied.
- Autonomy levels (docs) — None / Basic / Plus / Semi-Auto / Full-Auto control auto-context, auto-apply, auto-exec, auto-debug.
- Versioned plans and branches → replay-time-travel (docs) — Every plan update is versioned; branches explore multiple paths or models.
- Project map (tree-sitter) → agentic-rag (docs) — Fast project map generation and syntax validation across 30+ languages.
- REPL (docs) — REPL with fuzzy auto-complete for commands and file loading.
Patterns this full-code implements —
- ★★Approval Queue
Tentative-diff review is the central UX; nothing touches project files until accepted.
- ★Code-as-Action Agent
Diffs and shell commands are the primitive actions, staged before application.
- ★★Plan-and-Execute
Plan is the central first-class object - generated, branched, versioned, then executed.
- ★★Replay / Time-Travel
Plan-level version history plus branches for exploring multiple paths or models.
- ★★Code Execution
Plandex runs commands and auto-debugs in full-auto; controlled rollback is integral.
- ★★Multi-Model Routing
First-class multi-provider support with curated model packs trading off capability/cost/speed.
- ★★Agentic RAG
Tree-sitter project map is a graph-based retrieval substrate covering 30+ languages, not vector RAG.
- ★Todo-List-Driven Autonomous Agent
Plans function as checklists in spirit, but no explicit verbatim 'todo list' UI surfaced on the README/homepage; honest none.
Neighbourhood
Click any neighbour to follow the lineage. Scroll to zoom, drag to pan.