A new survey of the CLI coding-agent market, published July 6 by the arcbjorn blog, concludes that Anthropic’s Claude Code, OpenAI’s Codex CLI, and the open-source tool Omp now produce comparable output quality, so ranking them first, second, and third has stopped being a useful exercise. The practical stakes for engineering teams are direct: choosing a coding agent is no longer mainly a bet on which lab trained the smartest model. It is instead a bet on which harness surfaces the right tool, permission, and context at the exact moment the agent needs it.
That framing matters because the market the blog tracked has gotten crowded fast. The arcbjorn blog counted 35 CLI agents under active maintenance as of July 3, spanning lab-built tools, platform products from GitHub and JetBrains, and a long tail of open-source harnesses.
All three leading tools can read an unfamiliar repository, draft a plan, edit across multiple files, run tests, recover when a step fails, and hand back a patch shaped for production, according to the arcbjorn blog. What actually separates them, the blog argues, is task clarity, how clean the target repository already is, the permission model a team has configured, and whether the tool reaches for the right capability without being told.
Claude Code, the tool that defined the category when Anthropic shipped it as a research preview in February 2025, still has the cleanest interactive feel. Its planning loop reads as natural, and ambiguous, conversational work tends to land well there. The cost is lock-in: it runs only Claude models, stays closed source, and now bills autonomous, unattended usage against a separate credit allotment, a split Anthropic made in mid-June.
Codex CLI, OpenAI’s Rust-rewritten tool, is Apache-2.0 licensed with roughly 95,000 GitHub stars, and it adds operating-system-level sandboxing that neither Claude Code nor GitHub’s Copilot CLI currently offers. It gained browser control in April, a plugin marketplace, and a one-command import of Claude Code configuration, a feature that reads as OpenAI actively courting engineers who want to switch.
Omp, formally oh-my-pi, is the outlier. It is a community-maintained, feature-maximalist fork of the minimalist Pi agent, and the arcbjorn blog credits it with the widest feature set among the tools it examined closely. Omp ships hash-anchored edits, a project-reported 61 percent reduction in output tokens, built to stop patches from drifting onto the wrong text once whitespace has shifted. It also offers AST-aware rewrites across more than 50 language grammars, direct debugger control, and an Advisor: a second model that watches every turn and adds live corrections mid-stream. Its backing is thin next to Anthropic or OpenAI, a genuine hazard for a tool teams come to depend on daily, the blog notes.
OpenCode takes a different bet entirely. It carries the most GitHub stars of any tool in the survey, 182,000, and supports more than 75 model providers. It does not win on raw output quality. But because its maker earns money by staying neutral across models, it works harder than any other harness to perform well with whichever model a team already prefers, from GLM-5.2 to DeepSeek V4.
That pattern, a product category where the core agentic loop has become table stakes and competition has shifted to secondary features like editing precision and permission granularity, mirrors what happened to web browsers and mobile app stores once their underlying engines converged: differentiation moved up the stack, into surrounding tooling rather than the core function itself.
For a team picking a coding agent this quarter, the useful test is no longer a leaderboard. It is whether the harness’s permission model, memory, and tool access match how messy the target repository actually is, and whether an exit path stays open if a vendor changes terms, as Google did when it retired Gemini CLI’s free tier in June.
The arcbjorn blog published its mid-2026 survey of CLI coding agents on July 6, 2026.
