Overview
Celeste CLI is a standalone agentic development assistant for the terminal. It combines a Bubble Tea interface with autonomous tool use, project-aware context, code intelligence, session persistence, and a configurable persona.
The tool supports interactive conversations as well as longer agent runs, making it usable for both focused coding questions and multi-step development work.
Code Intelligence and Tools
Celeste CLI includes 48 built-in tools spanning file operations, shell tasks, web search, Git, code review, collections search, cryptography, and subagent orchestration. Its code graph combines MinHash and BM25 ranking, LSH-based queries, structural reranking, and tree-sitter parsing for TypeScript call-graph edges.
Graph-based review can detect stubs, placeholder implementations, lazy redirects, swallowed errors, and hardcoded values. Dedicated MCP tools expose indexing, code search, review, graph, and symbol operations directly from the cached graph without routing results through a chat model.
Agent Workflows
- Chat mode provides an interactive session with automatically continued tool calls and a 25-turn safety cap.
- Agent mode runs an autonomous multi-turn task with planning, file access, checkpoints, and resume support.
- Orchestrator mode adds a second reviewer model that critiques the agent run.
Project context can be discovered from its persona-themed configuration as well as common agent instruction files. Sessions are saved automatically in JSONL, while file checkpoints include stale detection and revert support.
Models and Interaction
The assistant works with several hosted model providers and local OpenAI-compatible servers. It also supports model reasoning controls and image input when the selected model has vision capabilities.
The terminal interface provides flicker-free rendering, live session cost tracking based on per-model pricing, and a persistent project experience. Hooks can run around tool calls, prompts, and sessions, with repository-defined hooks requiring approval before execution.
Permissions and Safety
A layered permission system applies allow, deny, or ask rules through pattern matching. The autonomous workflow also uses turn limits and checkpoints, giving users controls around extended tasks while retaining the ability to resume or revert work.