← Back to Reviews | Development

MicroCodex Review 2026 — A Sub-1MB C++ Coding Agent for Your Terminal

Marcus Webb · · Rated 6.8/10 · Free and open source (Apache-2.0). Bring your own ChatGPT/Codex plan via OAuth; the install is a single curl pipe
6.8 / 10
Ease of Use 8
Features 5
Value for Money 8
Performance 8
Support & Ecosystem 5

✅ Pros

  • • Sub-1MB binary that runs instantly on Mac and Linux — no Node runtime, no Electron, no 100MB+ install; just a single C++23 executable
  • • Uses your existing ChatGPT/Codex plan via OAuth (microcodex login), so there's no separate subscription or API-key management
  • • One-shot prompts and an interactive terminal UI for the same tool — good for both scripts and hands-on sessions
  • • Durable conversations with automatic context compaction, so long sessions don't blow past the context window without warning
  • • Discovers the same ~/.codex/skills filesystem skills as Codex, meaning existing skill sets port over directly
  • • Includes a bash safety denylist (blocks rm -rf, git reset --hard, forced git clean, disk-formatting, shutdown commands) before handing over the shell

⚠️ Cons

  • • Not feature-complete vs OpenAI Codex: no MCP support yet, text can't be copied from the terminal while running, and the denylist is a lexical guard, not a sandbox
  • • OAuth is tied to ChatGPT/Codex accounts — if you want to use other model providers, this isn't the tool
  • • Very young project: 12 GitHub stars, 0 forks, a single maintainer, and 44 commits as of August 3, 2026
  • • HN reception was lukewarm: commenters questioned whether small size matters and whether it has feature parity with larger harnesses (it doesn't, yet)
  • • No MCP support means no tool ecosystem integration — a hard miss for agent-heavy workflows in late 2026
Best For

Developers who want a fast, lightweight Codex-compatible coding agent in the terminal, especially on remote/headless machines, and anyone who wants to inspect a readable C++ implementation of an agent harness

Pricing

Free and open source (Apache-2.0). Bring your own ChatGPT/Codex plan via OAuth; the install is a single curl pipe

Quick Verdict

MicroCodex (github.com/paoloanzn/microcodex) is an ultra-lightweight coding agent written in C++23 that reimplements OpenAI’s Codex in a sub-1MB binary. It launched as a Show HN on August 2, 2026, and the pitch is contrarian: while every other agent harness is growing into Electron apps and 100MB+ installs, MicroCodex fits in less space than a single screenshot.

The core loop is genuinely nice — one-shot prompts, an interactive terminal UI, local coding tools, durable conversations with automatic context compaction, and OAuth into your existing ChatGPT/Codex plan. At 6.8/10, this is a Silver-tier review: the engineering is clean and the UX is refreshingly fast, but it’s a young project with real feature gaps (no MCP, no clipboard, no sandbox) and a skeptical HN reception that has a point.

What MicroCodex Actually Is

MicroCodex is a from-scratch C++23 reimplementation of the Codex agent harness — the same class of tool as OpenAI’s Codex CLI and Anthropic’s Claude Code, but built to be minimal. The repo (Apache-2.0, 44 commits) is organized into readable modules: agent.cpp, bash.cpp, edit.cpp, context-compaction.cpp, oauth.cpp, skills.cpp, http.cpp, and a termbox-based TUI. It is not a wrapper around Codex — it talks to the same backend endpoints via its own HTTP client and OAuth flow.

The headline spec:

  • Language: C++23, single binary
  • Size: under 1MB
  • Platforms: macOS (arm64 + x86_64), Linux (x86_64 + arm64)
  • Auth: OAuth with your ChatGPT/Codex plan (microcodex login, or --device-auth for headless)
  • License: Apache-2.0

Installing and Running It

Install is one line on Mac or Linux:

curl -fsSL https://github.com/paoloanzn/microcodex/releases/latest/download/install.sh | sh

Then sign in and start:

microcodex login        # opens browser; use --device-auth on headless boxes
microcodex              # interactive TUI session
microcodex "Find the failing test, fix it, and run the relevant test suite"   # one-shot

Credentials are stored under $CODEX_HOME (default ~/.codex), and MicroCodex discovers the same filesystem skills Codex uses (~/.codex/skills), so any skills you’ve installed for Codex — with a SKILL.md frontmatter of name + description — are picked up automatically. Skill metadata is added at session start; the full skill text is read only when its name or description matches the task, keeping the context window lean.

Building from source is also straightforward:

git clone --recurse-submodules https://github.com/paoloanzn/microcodex.git
cd microcodex && make
# binary lands at build/microcodex; tests via make test

That requires a C++23 compiler, make, and (on Linux) libcurl and OpenSSL dev files.

Feature Walkthrough

One-Shot Prompts + Interactive TUI

Two modes from one binary. The one-shot mode (microcodex "fix the failing test") fits scripting and CI; the interactive mode gives a terminal UI with shell highlighting and session history. In practice the TUI feels like a stripped-down Codex/Claude Code — arrow-key navigation, editable prompts, streaming responses.

Durable Conversations + Automatic Context Compaction

Conversations persist across sessions, and when the context fills up, MicroCodex compacts automatically — it summarizes older turns rather than hard-truncating. That’s a genuinely useful behavior for long agent sessions, and one many larger tools still handle badly.

Local Coding Tools

The agent gets filesystem tools (read, write, edit, glob) plus a bash tool. The bash tool is where the safety design lives: before starting your shell it applies a lexical denylist that blocks rm -f/rm -rf, git reset --hard, forced git clean, git checkout --, disk-formatting tools, and shutdown commands.

The README is refreshingly honest about the limits:

“This guard is not a shell parser and is not a complete security boundary: commands that do not match the denylist and all file operations run with the same permissions as the MicroCodex process.”

So treat it as a speed bump for accidents, not a sandbox.

Known Limitations (from the README)

  • MCP support is not implemented yet — no tool ecosystem integration, a significant gap for agent-heavy workflows in 2026
  • Text cannot be copied from the terminal while using MicroCodex
  • The bash safety gate is a lexical denylist, not a sandbox — indirect or unrecognized destructive commands may not be blocked

The HN Reception (Honest Signal)

The Show HN thread was skeptical, and the pushback is worth taking seriously:

  • fractorial: “I’m not quite sure why the harness itself needs to be small. Isn’t the system prompt and management of system prompt the bit you want lightweight?”
  • selcuka: “A harness needs to be good. Being large is annoying, sure, but it’s generally not an issue. Unless this one has feature parity with the ‘large’ ones (hint: it doesn’t), I don’t see any reason to use it only because it’s small.”
  • K0IN: “I think it’s nice to have options, but I also don’t see the appeal in ‘sub x mb’ (except as codegolf or just for the hacky part…).”
  • userbinator, in the project’s defense: “Why does it need to be large? should be the question instead.”

The fair summary: smallness alone isn’t a feature, but there IS a real use case for a codebase you can read end-to-end, a binary that starts instantly, and an agent that works on a headless box with --device-auth. The counter-argument — you’re missing MCP and ecosystem parity — is also correct, today.

Who Should Use It

Good fit:

  • SSH-box developers who want a coding agent on a remote/headless server (--device-auth)
  • People who want to read and fork a complete, understandable agent harness in one afternoon
  • Budget-conscious users already paying for ChatGPT/Codex who want a zero-additional-cost CLI
  • Anyone tired of 100MB+ agent installs

Skip it if:

  • You need MCP integration or a large tool ecosystem
  • You want model-provider flexibility (it’s ChatGPT/Codex-account-bound)
  • You need a real sandbox for untrusted code execution

Alternatives

ToolSizeModel accessMCPNotes
MicroCodex<1MBChatGPT/Codex (OAuth)No (planned)Fastest startup; readable C++ source
OpenAI Codex CLI50-100MB+OpenAI/ChatGPTYesFull feature set, heavier
Claude Code50-100MB+AnthropicYesBest-in-class agent UX, Node-based
Cline (VS Code)ExtensionAny providerYesIDE-native, not terminal

Verdict

MicroCodex is a clean, honest piece of engineering: a sub-1MB C++23 agent harness that runs instantly, reuses your Codex subscription and skills, and compacts context intelligently. It is not yet a Codex replacement — no MCP, no clipboard, no sandbox, and a 12-star community. But for a specific niche — fast terminal agents on headless machines, and a codebase you can actually read — it’s the most approachable option on the market. Watch it: if MCP lands and the star count climbs, this becomes a serious lightweight contender.

Rating: 6.8/10. Silver tier. Best for remote-box developers and open-source-curious agent users; wait for MCP before making it your daily driver.

MicroCodex Codex Coding-Agent OpenAI C++ CLI Open-Source AI-Coding Terminal