23 open source tools compared. Sorted by stars. Scroll down for our analysis.
By Erik Loyd, SaaS CEO and former COO/CFO of an AWS Premier Partner.
| Tool | Stars | Velocity | Score |
|---|---|---|---|
cc-switch A cross-platform desktop All-in-One assistant tool for Claude Code, Codex, OpenCode, openclaw & Gemini CLI. | 154.8k | +1866/wk | 91 |
caveman 🪨 why use many token when few token do trick. Claude Code skill that cuts 65% of tokens by talking like caveman | 120.6k | +2099/wk | 91 |
oh-my-claudecode Teams-first Multi-agent orchestration for Claude Code | 40.9k | +149/wk | 90 |
magpie Every agent's model. One place. Codex on DeepSeek, Claude Code on Kimi, from the menu bar. | 37.3k | +4565/wk | 86 |
DSH Desktop A community-built desktop app for DeepSeek Harness, with the plugin market and mobile remote control built in. Free, MIT, and not an official DeepSeek project. | 36.1k | +662/wk | 88 |
codex-plugin-cc Use Codex from Claude Code to review code or delegate tasks. | 35.2k | +187/wk | 73 |
dsh-market The plugin market inside DeepSeek Harness: browse, search, one-click install · DSH 可视化插件市场 | 12.6k | +977/wk | 82 |
codeburn See where your AI coding tokens go. Interactive TUI dashboard for Claude Code, Codex, and Cursor cost observability. | 11.8k | +91/wk | 86 |
dsh-web Plugin and skin collection for DeepSeek Harness (DSH) Web UI - task board, git graph, right-side panel, remote mobile UI, pet, live token stats, and skin center. | 10.4k | +179/wk | 70 |
html-anything ✨ The agentic HTML editor: your local AI agent writes the HTML, you ship it. 🚀 75 Skills × 9 Surfaces (magazine · deck · poster · XHS / tweet · prototype · data report · Hyperframes) 🛡️ Sandboxed preview · 📤 1-click to WeChat / X / Zhihu / HTML / PNG 🔑 Zero API key. Claude Code / Cursor / Codex / Gemini / Copilot / OpenCode / Qwen / Aider. | 10.0k | +557/wk | 84 |
PI-Desktop Local-first AI coding agent desktop: Electron + Rust host core + pi Agent Harness + user-installable plugins | 8.0k | +278/wk | 73 |
dsh-TUI DSH 官方公众号收录的 TUI 补位插件:Claude Code 风,鲸鱼顶栏/实时状态/流式思考/双击 Esc 回滚/上下文进度+TPS。npm 一键装。 DSH official WeChat featured TUI plugin, Claude Code style: whale bar, live status, streaming thoughts, double-Esc rollback, context bar + TPS. npm one-click. | 7.8k | +498/wk | 78 |
oh-my-hermes ⚚ The coding harness with Hermes Agent wired into IDE. | 5.7k | +477/wk | 76 |
earshot Ears for coding agents: let AI hear, review and make sound. CPU-only, open source. | 4.9k | +643/wk | 76 |
vibe-wise A Claude Code plugin that helps you learn how to build while AI writes the code. | 4.4k | +126/wk | 62 |
codex-plusplus Codex++ tweak system for the Codex desktop app | 3.8k | - | 62 |
jevgrep Find code by asking what it does. A CLI for coding agents that uses Jev to discover relevant files and source context. | 3.2k | +26/wk | 72 |
codex-host Run Pi and Claude Code directly in Codex Desktop. 在 Codex Desktop 中直接运行 Pi 和 Claude Code。 | 3.2k | +68/wk | 72 |
claude-usage A local dashboard for tracking your Claude Code token usage, costs, and session history. Pro and Max subscribers get a progress bar. This gives you the full picture. | 2.3k | +22/wk | 59 |
deepclaude Use Claude Code's autonomous agent loop with DeepSeek V4 Pro, OpenRouter, or any Anthropic-compatible backend. Same UX, 17x cheaper. | 2.3k | +1/wk | 59 |
keep-codex-fast A backup-first Codex skill for keeping local Codex state fast, clean, and recoverable. | 1.6k | +8/wk | 47 |
mindwalk A visualization tool that replays coding-agent sessions on a 3D map of your codebase. | 1.4k | +3/wk | 53 |
codex-shim Local Responses-API shim that exposes Factory BYOK models (and optional ChatGPT GPT-5.5 passthrough) to Codex Desktop. | 1.1k | - | 60 |
Stay ahead of the category
New tools and momentum shifts, every Wednesday.
CC Switch is a desktop app for juggling API providers across your AI coding tools. Instead of hand-editing JSON, TOML or YAML, you pick a provider from 90+ presets (official logins, AWS Bedrock, NVIDIA NIM, community relays), enter a key and switch in one click. It also manages MCP servers, skills and prompts across Claude Code, Claude Desktop, Codex, Gemini CLI, OpenCode, Grok Build and a handful of others. MIT licensed and free. Install from the releases page, Homebrew or AUR; it's a Tauri app for Windows, macOS (signed and notarized) and Linux. It edits only the key fields in each tool's config and backs up the original on first write, so uninstalling leaves your tools working. An optional local proxy on 127.0.0.1 handles format conversion and automatic failover between providers, and settings can sync over WebDAV or S3. Solo devs bouncing between official plans, Bedrock and cheaper relays get the most here. Teams that standardize on one provider don't need it. A beta OAuth center lets you use ChatGPT, Copilot and Grok subscriptions as providers inside Claude Code and Codex; the README itself warns that may violate those vendors' terms. The catch: it's funded by a wall of API relay sponsors advertising steep discounts on official model prices, and relays show up among the built-in presets. Routing your code through a discount reseller is a trust decision, not a config detail. Make it on purpose.
Caveman makes your coding agent say less and read less. It began as a joke skill that made Claude talk like a caveman. Version 4 is a local proxy plus skills for Claude Code, Codex, Gemini CLI and 30+ other agents: shorter replies, compressed logs and JSON before the model reads them, and saved scripts the agent would otherwise rewrite. The whole repo is Apache 2.0 and free.
npx @caveman-ai/cli@latest finds your agents, shows what it will change, and applies on Enter (Node.js 22.13+). Want only the voice? npx skills add JuliusBrussee/caveman -g. The CLI sends usage stats by default, never prompts or code; caveman telemetry off stops it.
The paid part is Caveman Cloud, for teams building their own agents. Free covers one person; Team is $300/month with $300 of usage included. Solo and small teams: free. Teams shipping agents: Cloud.
The catch is the headline number. The repo still advertises 65%, but against a plain "Answer concisely" prompt, the README's own table has the default mode cutting about 3% more output. The bigger win is input compression.
oh-my-claudecode turns one Claude Code session into a small dev team. Type /autopilot "build a REST API" or /team 3:executor "fix all TypeScript errors" and it plans, builds, verifies and loops on fixes with specialized agents, no prompting every step. It's a free, MIT-licensed plugin from the Claude Code marketplace, also published on npm as oh-my-claude-sisyphus (same author, odd name).
Nothing to host. It runs inside your session and keeps state, logs and transcripts in a local .omc/ folder. From the terminal, omc team opens tmux panes running Codex, Gemini, Grok, Cursor or Claude CLIs, so a second model can review the first. There is no paid tier, but you need a Claude Pro or Max plan or an Anthropic API key, plus accounts for any other CLI you call. Solo developers and small teams get the most from it; larger teams should pin a version before standardizing.
The catch: it ships up to three releases a day and drops features along the way (swarm became team, the Codex and Gemini MCP servers went in v4.4), so older setups break. Parallel agents chew through usage limits fast. Plain Claude Code subagents cover much of this with fewer moving parts.
Magpie is a menu-bar app that shows every coding agent on your machine and which model each one uses, and switches them with a click. Behind that sits a local gateway that translates between the OpenAI, Anthropic and Gemini APIs, so Codex can run on DeepSeek or Claude Code on Kimi. It covers 24 agents, around twenty provider presets and any compatible endpoint you add. MIT, free, from yetone, the developer behind avante.nvim. It's a Go app under 15MB for macOS, Windows and Linux, plus a terminal-only build. It can also hand your signed-in subscriptions (Claude, ChatGPT, Copilot, Gemini) to other agents as providers, and rotate across several accounts to use up each one's allowance. Mac builds are signed; Windows and Linux builds aren't yet. Developers juggling several agents and API keys who want one switchboard: try it. For Claude Code routing alone, claude-code-router is simpler, and BerriAI/litellm is the self-hosted server version of the gateway. The catch is the subscription routing. The README says it drives the real Claude binary so other agents' traffic stays out of Anthropic's third-party detection, and warns that Google may suspend Antigravity accounts used outside Antigravity. That's working around your providers' terms. Stick to API keys, or use an account you can afford to lose.
DSH Desktop turns DeepSeek Harness into a normal app. The harness is DeepSeek's open source agent runtime, and running it means Node, pnpm, a terminal, and a port. This is one installer for Windows x64 or any Mac, Intel or Apple Silicon, and the Web UI, host service, and plugin system open in a native window. MIT, free, and not an official DeepSeek project. Nothing to host. The installer bundles Electron, Node, pnpm, and a pinned harness build, and leaves your global PATH alone. Updates download only after you confirm. Upstream runs unmodified, so official plugins install from the bundled terminal. Newer builds add a built-in community plugin market and mobile remote control. Solo and small teams on Windows or a Mac should use this instead of the npx command. Medium and large teams get nothing extra: no shared profiles, no policy controls. Anyone already running the harness in a browser gets more from zhu1090093659/dsh-web-ui. For a multi-provider chat client instead, CherryHQ/cherry-studio is further along. The catch is what it pins. Upstream calls itself a developer preview and warns in capitals that breaking changes are coming, and this ships a fixed release candidate. No Linux installer has shipped yet, the issue tracker carries launch crashes on both supported platforms, and LAN mode opens the UI to your network with no authentication.
codex-plugin-cc is OpenAI's official plugin for running Codex inside Claude Code. It adds a read-only Codex review, a steerable "adversarial" review that challenges your design, and /codex:rescue to hand Codex a bug in the background. The plugin is free under Apache 2.0.
Setup is four commands in Claude Code, ending with /codex:setup, which checks for the Codex CLI and offers to install it. It uses your existing Codex login and config. /codex:transfer turns a Claude Code session into a Codex thread you can resume in Codex.
The real cost is Codex usage: you sign in with a ChatGPT account or an OpenAI API key, and every review counts against your Codex limits. The README says a free ChatGPT account works, but OpenAI pitches Free and Go ($8/month) at quick tasks; Plus at $20/month is the first plan listed with Codex in the CLI. Solo and small teams: Plus per developer. Larger teams: Business or Enterprise for SSO.
The catch: you're paying for two agents. Claude Code needs its own plan, and the optional review gate runs a Codex review every time Claude tries to stop, which OpenAI warns may drain usage limits quickly. No commits since early July.
dsh-market is the plugin store for DeepSeek Harness, DeepSeek's open source agent harness, and several desktop builds of the harness ship it by default. Browse the community catalog with screenshots and comments, then install, update, disable or remove plugins with a click. MIT licensed, free.
Setup is one command, dsh plugin --profile web add dshmarket, and a restart. It has grown well past a store: themes, plugin groups, backups to WebDAV or a private Gist, load-order diagnostics, and a recovery screen when one bad plugin stops the harness from booting.
No paid tier, no telemetry. Solo and small teams: free. Running dsh under pm2 or launchd? Set allowRestart to false in the profile config; only systemd is detected automatically.
The catch is trust and churn. The store installs only from one curated registry, and the same maintainer runs both, so one person decides what reaches your harness. The project says plainly that listing is not endorsement: read a plugin before installing it. The harness itself is a developer preview that promises breaking changes, and dsh 0.2.1 added its own Discover tab that searches this same catalog.
Codeburn shows you where your AI coding tokens go. Install it, run it, and get a TUI dashboard breaking down cost by project, model, and session across Claude Code, Codex, Cursor, Copilot, and others. No API keys needed: it reads the session files these tools already store on your machine. Setup is one command: npx codeburn. Pricing data pulls from LiteLLM and caches for 24 hours. Beyond raw cost, it classifies your sessions into task types (debugging, testing, coding) and calculates one-shot success rates so you can see which tasks burn tokens on retries. Anyone paying for AI coding tools should run this once just to see the numbers. The macOS menubar app gives passive cost awareness without opening a terminal. The optimize command flags waste patterns like repeated failures on the same task. The catch: it only knows about tools that store session data locally. If your AI tool does not write to disk (or you have not used it on this machine), codeburn cannot see it.
dsh-web turns DeepSeek Harness's browser UI into a workbench. It's a plugin pack for dsh web and the DeepSeek Harness desktop app: a task board that runs real agent sessions on cron schedules, a git graph that gives each session its own worktree so parallel agents don't trample each other, an SSH panel, token and balance tracking, and a phone remote you pair by QR code. Apache 2.0 and free.
Install is one command, dsh plugin --profile web add @linxin666/dsh-web-all@latest, or paste it into the desktop app's plugin page. Remote access is one switch: the auto tunnel ships cloudflared. Turn it on and a stable-hostname relay is on by default, a Cloudflare Worker run by the package author, so your agent traffic crosses his infrastructure. Anonymous install telemetry is also on by default.
No paid tier; the only bill is the model. Solo and small teams: free, and worktree isolation alone earns the slot. Larger teams: there are no user accounts or roles, just device pairing.
The catch: it rides DSH's plugin API, still on release candidates. Some skins also use art the project doesn't own, including backgrounds extracted from the War Thunder game client, and several are non-commercial only.
HTML Anything turns Markdown, spreadsheets, PDFs or plain notes into polished single-file web pages by handing them to the coding agent you already use. Pick Claude Code, Codex, Gemini or one of about twenty supported CLIs, pick a template, and it streams back a slide deck, magazine layout, poster or social card. Apache 2.0, from the team behind nexu-io/open-design, and it reuses your logged-in agent session, so there's no API key.
It runs as a local web app in your browser, not a desktop program. Clone it, install with pnpm, and start it with pnpm -F @html-anything/next dev (plain pnpm dev does nothing at the repo root). The 81 templates are agent skill files, and output exports to WeChat, Zhihu, Notion, Bluesky, PNG or straight to your own Vercel account. A command-line version handles batch jobs.
Anyone already paying for an agent subscription who wants presentable HTML without another tool bill should try it. Without an agent CLI installed, it does nothing. The templates are Chinese-first, with CJK font stacks throughout, which suits WeChat publishing better than a US pitch deck.
The catch: to run unattended, it launches your agent with every safety prompt switched off (Claude Code in bypass-permissions mode, Gemini in yolo mode). That's an agent with full access to your machine working from whatever file you drop in, so don't feed it documents you didn't write. Development has also slowed sharply since spring, and there has never been a tagged release.
PI-Desktop is a desktop app for AI coding agents: open a local repo, connect a model, and the agent reads code, edits files, runs commands and shows you the diff. Under the hood it runs pi, Mario Zechner's open source coding agent, wrapped in persistent sessions, a plan mode, subagents and a plugin system. It's free under LGPL-3.0, with no account. Setup is a download for macOS, Windows or Linux. You bring the model: OpenAI, Anthropic, any OpenAI-compatible endpoint, or local models through Ollama or LM Studio. Requests go straight to your provider, keys sit in the OS keychain, and the app has no telemetry. It can also import your Claude Code, Codex and OpenCode sessions. There's no paid tier; your only bill is the model. Solo: free, and a good GUI if terminal agents aren't your thing. Small teams: free, each person brings their own keys. Larger teams: no documented admin, SSO or central billing, so that's still Cursor or Copilot territory. The catch: it's an early preview shipping three releases in a single day, so expect breakage. The author also publishes an optional plugin listing free community API relays. Don't send your code through an anonymous relay.
dsh-TUI gives DeepSeek Harness the terminal interface it doesn't ship with. It's a Claude Code style full-screen front end for DeepSeek's open source agent runtime: streaming markdown, a context meter, tokens per second, a running cost estimate, and double Escape to rewind a turn. MIT licensed and free.
Setup is one line, npm install -g @deepseek-ai/dsh @deepseek-harness-tui/dsh-tui, then dsh-tui. You need Node 22.19+, pnpm, and a DeepSeek key or account. It now also drives Claude Code and Codex as experimental backends, so one terminal UI can sit on three agents. Change two defaults first: on Windows the profile gives tools full access with no approval prompts, and DSH attaches incremental session logs to requests sent to your model endpoint.
The plugin is free; the model isn't. DeepSeek V4-Pro bills $1.32 per million input tokens and $3.96 per million output at peak, half that off-peak, and Flash is cheaper. Solo: free, worth an afternoon. Small teams: free, as a personal preference. Larger teams: don't standardize on it.
The catch is churn. The harness is still on release candidates and the TUI ships a new pre-1.0 release every few days. Claude Code, Codex and OpenCode can all point at DeepSeek's API directly.
oh-my-hermes is an add-on for Nous Research's Hermes Agent (NousResearch/hermes-agent) that splits one goal into parallel lanes, each with its own model and skill. Type /dag-ulw and your request becomes a graph of small tasks; ready ones run at once, and a separate node checks the goal before you get a report. MIT licensed, free, independent of Nous. Install is a curl script, Homebrew, npm or Bun, then omh setup, and you need Hermes working first. Each work category (deep, quick, visual-engineering and more) has an editable model chain, and when a provider rejects a model it moves down the chain instead of quietly downgrading. A beta desktop app, omher, puts Hermes bots next to Claude Code and Codex sessions. Solo: free, and worth it if you already live in Hermes and pay several model providers. Small teams: free; share one set of model chains. Larger teams: no team controls. Claude Code users get the same idea from Yeachan-Heo/oh-my-claudecode; code-yeongyu/oh-my-openagent, where ultrawork started, is now its own agent. The catch: lanes multiply token spend, and the shipped chains lead with premium models. Check your chains before the first big graph run.
earshot gives coding agents ears. An agent can build a product video in code but can't hear that the whoosh lands late or the key click sounds cheap. earshot turns audio into text an agent can reason about: what a file sounds like, how well it matches a description, where the best moment sits in a long recording, and a timestamped critique of a whole film. MIT licensed and free.
It runs on CPU as a CLI, a Python API and an MCP server, plus two agent skills. The first run downloads about 1.1 GB of open models, then it works offline. Each CLI call reloads the models in 20 to 40 seconds, so use the MCP server for real work. Version 0.3 adds 201 sound effects generated from code.
No paid tier; the only paid piece is an optional Gemini critique on your own key behind an explicit upload flag. Solo, small teams and larger teams: free, installed per machine.
The catch: it's brand new, one author, alpha, and the README admits its sound search can game its own models. And don't pip install earshot: that PyPI name belongs to an unrelated voice-detection library. Install from the repo.
VibeWise turns Claude Code or Codex into a pair-programming tutor. You propose the design, the agent asks questions and explains unfamiliar concepts at named checkpoints, then writes the code you agreed on. Its tagline is "You build. AI writes." It's a free MIT plugin for people who want to learn while they ship, from career-changers to experienced engineers in an unfamiliar stack.
In Claude Code it now installs from Anthropic's plugin directory (/plugin install vibe-wise@anthropic-plugin-directory); restart and run /vibe-wise:learn. Codex installs from the author's GitHub marketplace and has no hook, so you rerun $vibe-wise:learn each session. Notes live in a .vibe-wise/ folder per project. No account, backend or telemetry.
Builders without a CS background who code with AI and want to understand what they're shipping get the most here. Anyone who just wants output should skip it: it slows the build down on purpose. Anthropic's free learning-output-style plugin is the lighter option.
The catch: one maintainer, no tagged releases, and ports to OpenCode, Cursor and Copilot CLI still sit in unmerged PRs. Install from the directory, which reviews each version. The GitHub route with auto-update on lands hook changes you never looked at.
Codex++ was a tweak loader for OpenAI's Codex desktop app. It patched your local install so Codex would load community-written tweaks at startup: custom settings pages, UI changes, main-process code, even native macOS APIs through its own bridge. MIT licensed, free, macOS, Windows, and Linux. Past tense on purpose. The maintainer archived the repository on August 15, 2026, and it is now read-only. Version 1.0.0 was real work, with a Homebrew formula, a doctor command for signature and permission failures, a safe mode that disables tweaks without deleting them, and detection for Owl, the native shell current macOS Codex builds use. None of it gets another commit. Do not install this now. The whole design depends on somebody tracking OpenAI's app changes and shipping a patch when Codex breaks the loader, and nobody is doing that anymore. For an AI coding tool you can actually modify, use one that is open to begin with: anomalyco/opencode, cline/cline, and Aider-AI/aider all take extensions without touching a signed app bundle. Solo, small, or large, the answer is the same. The catch was always that this was unofficial, patched a signed application, and voided its code signature. Archiving turns that from a tolerable risk into a dead end. If it is already installed and working, run codexplusplus uninstall to restore the original app while the tooling still runs.
jevgrep lets a coding agent find code by asking what it does instead of grepping for names. Ask where the app retries failed webhooks and it returns the relevant files plus source excerpts for the agent to read. MIT, a Node CLI plus an agent skill. It doesn't run locally. Every search sends your source to Jev, TypeSafe's hosted classification model, through Vercel's AI Gateway, OpenRouter or TypeSafe directly. Jev costs $42 per billion input tokens with output free, so the search bill itself is tiny. Declaration parsing covers Python and TypeScript/JavaScript; everything else gets a plain-text fallback. The pitch is cheaper agent runs. On ten SWE-bench tasks the author reports the same 8 of 10 solved at 25.8% lower total cost, Jev included. Promising, but ten tasks is a small sample. Agents working in big unfamiliar repos: try it. Small repos: the README itself says ripgrep may be all you need. The catch is where your code goes. Every query uploads source to a third-party model. For private or regulated code, that's a call for your security team, not your agent.
CodexHost runs other coding agents inside Codex Desktop. If you have settled on one desktop app but keep dropping to a terminal for Claude Code or Pi, this puts those harnesses behind the Codex UI instead. It supports Pi, Claude Code, OpenCode, DeepSeek Harness, Grok, Oh My Pi, and Antigravity, and it can hand a task from one agent to another. The interesting call is that it does not flatten everything into one protocol. Each agent goes through its own native interface, Pi over its official RPC and Claude Code through the Agent SDK, with a CLI shim in front. Streaming, tool calls, diffs, and approval prompts come through the way the underlying agent produced them instead of as a lowest common denominator. Install is a global npm package, or an installer for macOS and Windows, with Linux on x64 and ARM64. MIT and free. Nothing to host. The agents themselves still bill you on their own plans. The catch is that this is glue over seven moving targets, and it needs official Codex Desktop plus Node 22.19 or newer. Every one of those vendors ships breaking changes on its own schedule. The question is not whether it works today. It is whether the author is still patching it in six months.
Claude Usage reads the JSONL session logs that Claude Code writes to your machine and turns them into charts. Per-model token breakdowns, cache hit rates, cost estimates, and session history, all in a local browser dashboard. Anthropic's own UI gives you a progress bar. This gives you the full picture. Zero dependencies. Standard library Python only, no pip install. Clone the repo, run the dashboard command, and it serves a Chart.js UI at localhost:8080 that auto-refreshes every 30 seconds. A SQLite database at ~/.claude/usage.db caches the parsed data for fast incremental re-scans. CLI commands cover scan, today, stats, and dashboard. Pro and Max subscribers who want to understand their actual token consumption per session and per model need this. The cost estimates use current API pricing, which is useful even for subscription users as a proxy for understanding usage patterns. The catch: only captures local Claude Code sessions. Cowork sessions (server-side) are not included. Cost estimates reflect API pricing, not what you actually pay on a Pro/Max subscription. Single-maintainer project, so pricing tables need manual updates when Anthropic changes models.
deepclaude swaps the model behind Claude Code without changing the interface. You keep the autonomous agent loop, file editing, and bash execution, and route calls to DeepSeek V4 Pro, OpenRouter, Fireworks, or any Anthropic-compatible backend. Free, MIT. Setup is two minutes. npm install, set environment variables for your chosen provider, and the CLI behaves like Claude Code with a different brain inside. Multiple providers can be configured at once, and you switch between them at the env-var level. Solo: real money. DeepSeek V4 Pro at $0.87 per million output tokens versus Anthropic's $15 means a hobby project can stop bleeding cash. Small teams: depends entirely on whether your model of choice handles your codebase as well as Sonnet does. Larger teams: probably stay on Anthropic. A productivity hit on a senior engineer costs more than the API bill ever will. The catch: Claude Code is good because of Sonnet and Opus, not because of the loop around them. DeepSeek is solid but it is a different model with different blind spots. Test it on your real code before committing.
keep-codex-fast is a maintenance script for OpenAI's Codex CLI. It walks through accumulated chat history, terminal logs, and worktrees, then archives or prunes them safely. MIT. The flow is conservative on purpose. Inspect first, write a handoff doc, back up state, then optionally apply changes. Read-only by default, archive instead of delete, and you opt in with --apply when you actually want it to do something. Solo: useful if your Codex sessions feel sluggish from months of chat history, terminal logs, and stale worktrees piling up. Teams: skip, this is personal-machine maintenance with nothing to centralize. The catch is the scope. This is for OpenAI's Codex, not Anthropic's Claude Code. The names sound similar, the tools are different, and the state directories do not overlap. If you run Claude Code, this script will not help you.
Mindwalk replays what your coding agent actually did. It renders your repo as a 3D map and plays back the session as light moving through it: which files the agent searched, read, and edited, in order. The diff tells you what changed; this shows you how the agent got there. MIT licensed and free. Install is a one-line script and it runs entirely on your machine. No server, no account. The only outbound call is optional: it can send a session summary to your own Claude Code or Codex account for evaluation. Worth a look if you run coding agents daily and keep wondering why the agent read half the codebase to change one file. It comes from the maintainer of air, the Go live-reload tool, which counts for something on follow-through. Free at any team size. The catch is that it's young and narrow. Version 0.2, small codebase, one use case. Tracing platforms like Langfuse cover LLM apps broadly; nothing else does this specific replay trick, but 'nothing else does it' cuts both ways.
codex-shim is a local proxy that lets Codex Desktop talk to models it was never built to support. OpenAI's Codex app normally locks you into a short list of models. This shim sits between the app and whatever provider you want, translating Codex's Responses API format into calls to Anthropic, DeepSeek, or any generic chat-completion endpoint, with an optional ChatGPT passthrough that reuses your existing subscription. MIT licensed and completely free. Running it means Python 3.11 or newer, a pip install, and editing a JSON config to point Codex at localhost. On macOS you may also patch the app so your custom models show up in the picker. None of this is hard for someone comfortable in a terminal, but it is fiddly, and it can break when Codex ships an update that changes its internal API. This is a workaround, not a product. Solo developers who want Claude or DeepSeek inside Codex without paying for a second tool get real value here. Small teams might standardize on it to route through a cheaper model. Larger teams should think hard before depending on an unofficial shim that lives or dies by one maintainer. The catch is fragility. You are patching around a closed app's limits, so every Codex release is a chance for this to stop working. Use it knowing that.