cloi: Run a Local Coding Agent in the Terminal

Overview

Cloi is a local-first coding agent for working with code through a terminal conversation. It can search and read a workspace, make changes, and run commands to help answer coding requests, with Ollama providing the models.

It is aimed at developers who want an agent to operate on local files while keeping data on their machine. Its approach emphasizes checks and permissions around tool use, rather than treating model output as automatically trustworthy.

Key Features

• Uses Ollama models to handle coding requests from an interactive terminal session or a single command.
• Selects primary and fallback models based on measured machine resources and model capability.
• Escalates to a larger fallback model when signals indicate that a turn is failing.
• Checks file-related claims, citations, and absence conclusions against workspace contents.
• Saves substantial tool results for later recall, including across conversations.
• Provides a persistent Python interpreter for combining stored results and calling tools.
• Summarizes older conversation history to manage long prompts.
• Contains workspace paths, requests permission for writes and commands, and sanitizes subprocess environments.

Use Cases

• A developer debugging a failing test can ask Cloi to inspect relevant files, run tests, and make a proposed fix.
• Someone navigating an unfamiliar repository can use it to search files and gather results without switching between terminal tools manually.
• Developers who want AI-assisted edits but do not want workspace data sent to an external API can run the workflow locally with Ollama.
• A user with repetitive inspection or per-file analysis tasks can combine tool results in Cloi's persistent Python session.

Getting Started

Install the package and launch Cloi:

npm install -g @cloi-ai/cloi
cloi

Cloi requires Ollama, Node 22.5+, and a tool-capable model. See the repository README for setup and usage details.

Alternatives

• AgentAleph: AgentAleph also edits projects with a local coding agent, but adds a desktop interface and GGUF model management instead of using Ollama in a terminal.

Considerations

• Ollama is the only supported provider; the README lists no web access, MCP, or sub-agents.
• Local operation still depends on having suitable hardware and a model that fits. Setup selects models based on available resources.
• The project documents a ceiling for difficult cross-file debugging: models that fit in 8 GB did not solve the tested hard cases. This is a stated limitation, not a guarantee about every task.
• Commands, writes, and edits require permission, so unattended workflows may need user interaction.