AI Coding Agents

I have a love-hate relationship with AI.

I have zero patience for overblown LinkedIn posts, probably written by AI itself.

The sound of 'vibe coding' makes me shiver.

I still secretly use AI to write bits of code and brainstorm.

Open Code is a 100% open-source, 0% affiliated terminal-based agent built by and for Neovim users.

Open Code is not to be confused with a discontinued project also called Open Code by another developer.

Open Code allows you to use any model, with an extensive list of supported models.

Open Code focuses on user experience with features like interface themes, autoloading LSPs, parallel models, and one-click session sharing with the team.

Open Code has an internal model router called Zen that finds the latest, most cost-efficient models using one payment, and Open Code doesn't profit from it.

Zen supports a pay-as-you-go model.

Zen is a model router with models tested and approved by the Open Code team, ensuring the latest updates automatically, with no profit for Open Code and credit card payment at provider cost plus processing fees.

I've been paying Cursor $20/month for six months probably not using 80% of it.

Cloud Code costs $17/month with no pay-as-you-go option.

With Zen, I only pay for what I use.

Open Code runs a local server, which is critical for accessing files, unlike cloud-based tools like Devon or Codec.

An agent is a loop talking to an LLM and iterating over a task until it requires intervention or completes all steps; it equals LLM plus tools plus loops.

Agentic mode involves providing permissions completely or manually, and the main risk is the limited context window.

Open Code has a solution for the limited context window, similar to solutions in Codex and Claude.

Open Code is open source on GitHub and super popular.

Installation can be done by curling a script.

The default theme is not integrated into the terminal window, but the /themes command shows many theme options.

The default model is Grock Code Fast, which is currently free because they are gathering training data.

The black boxes in the chat are the thinking steps yielded by the LLM, not final responses.

The user is actually conversing with Open Code, not with the underlying model Grock.

The /thinking command toggles the visibility of thinking blocks.

The /agents command shows a fuzzy-searchable list of agents.

Setup involves signing in, adding a credit card, creating an API key, and picking a provider.

There is an extensive list of onboarded providers beyond Zen; the team recommends either Zen or Claude directly.

Adding the API key and using /models lists available models through Zen; Sonnet 4.5 is currently the latest and greatest.

Grock is not present in benchmarks because they only benchmark Python, and Sweet Bench benchmark is only Python; no benchmark compares models on the same code base with same prompts.

To leave the UI, press Ctrl+C twice or type exit; opening Open Code with a project context requires providing the path after the command.

In less than a minute, Open Code can provide architecture, product goals, and tech stack overview from a codebase.

Every project should have an agents.md file to guide the AI agent with do's, don'ts, and instructions.

Open Code's /init command reads the project's files, conventions, methods, and even cursor rules and Copilot instructions to generate agents.md.

Default agents are 'build' (can modify files) and 'plan' (read-only), switchable with Tab.

The author recommends adding custom agents with specific instructions, temperature, verbosity, and dedicated model, configurable via markdown files rather than JSON.

Custom agent examples include a deep thinker using GPT-5 with high reasoning and low verbosity, and an email responder.

Open Code has a built-in list of LSP servers, but markdown is not one of them.

The default temperature setting is 0.1 (confident, finite), but can be increased to 0.8 for more creativity and randomness.

Custom commands can be created in Open Code for building, testing, git operations, and code reviews, by adding markdown files in the command directory.

A different model can be used for code review to simulate fresh eyes on changes.

To add an MCP, you must configure open_code.json, which holds keybindings, shortcuts, and other configurations.

The author integrates Sneak MCP by adding a simple command object in open_code.json and then can ask Open Code to scan the project for security, including dependencies and code.

The security scan found a low severity issue due to a hard-coded credential, but test files were not ignored by the scanner.

/sessions shows chat history with token count, context window percentage, and price paid.

/share creates a publicly accessible URL of the session, which can be unshared later.

/compact summarizes the conversation to free up context window, working well enough to feel like an infinite context window 90% of the time.

/export sends the session to your editor as a local file with a summary, requiring an editor environment variable.

Open Code lacks IDE integration compared to tools like Cursor and Windsurf, but opencode.nvim provides Neovim integration as a sub terminal.

With LazyVim, installing opencode.nvim adds keybindings: Leader O T to toggle, Leader A to ask about code at cursor, Leader O E to explain line, Leader O S for git review prompt.

The git review feature correctly warns if a change will break the code given the current config file.

Open Code runs a local server with a REST API that can be used to integrate the tool anywhere.

Open Code has a GitHub action integration that triggers an agent job when /oc or /open-code is mentioned in issues, using the chosen model to participate in the conversation.

Open Code is a pleasure to work with, focused solely on user experience, and ported into Neovim makes it the ideal setup.