Coding Agents for Multi-platform

The $200/month Max plan is what serious users need—the cheaper tiers hit limits fast on large codebases. Token consumption on big repositories gets expensive. There is no autocomplete; this is for delegating tasks, not for inline suggestions while you type.

You are switching editors—VS Code extensions mostly work, but not all. Cursor closes, your background agents stop. The context can drift on very large refactors; you will need to guide it back.

Devin is paid. The $500/month price prices out small teams. Success rates drop sharply on vague tickets; this is not a free coding agent.

The CLI is open source, but the models are not free: you pay through ChatGPT plan quotas or API spend. Codex Web runs on OpenAI’s cloud, not on your machine. The install.sh, brew and npm routes are supply-chain choices, so pin a release binary when that matters. Vs Claude Code or OpenCode: this is OpenAI’s first-party agent, tied to ChatGPT plans or OpenAI API keys.

Heavier than lighter agents like Goose. Self-hosting autonomous agents requires infrastructure and monitoring. The autonomy that makes it powerful also makes debugging failures harder.

Huge security surface. Every connected channel is a potential attack vector. You are trusting OpenClaw with access to your messaging platforms and the code it writes. Evaluate carefully for sensitive work.

Gemini-bound lock-in—not model-agnostic like Goose or Crush. Consumer access (free/Pro/Ultra Google login) ended 2026-06-18; enterprise and API keys still work. Don't build on the old Google-login quota path. The OSS CLI remains. Antigravity CLI is the successor, not a 1:1 swap. Preview/nightly channels move fast; pin stable for production workflows. Enterprise path pushes you toward Vertex AI and Google Cloud project setup.

Licensed FSL-1.1-MIT (competing-use restriction; source-available; MIT after two years—not OSI open-source). Pre-1.0 energy—config is migrating from JSON to crushrc Bash. Not a fully autonomous SWE agent like OpenHands; you stay in the loop. Hyper is Charm own model path if you want a subscription. Metrics on by default (DO_NOT_TRACK / CRUSH_DISABLE_METRICS to opt out).

On Windows, `wt` is Windows Terminal's command. The installer ships `git-wt` to avoid the collision (`winget install max-sixty.worktrunk` then `git-wt config shell install`); you can also disable the Terminal app-execution alias. Project hooks require first-run approval. This is worktree plumbing, not a coding agent.

GPL-3.0 is copyleft, not MIT. Hosted agent features that talk to Zed's models are a product on top of the editor — no Cloud free tier to count on. Web is not shipped.

Two catches worth reading on /log: “The date stamp is a trigger channel” (auto today’s-date × open-weight LoRA as a time-release fuse) and “Compaction ate the denial” (long sessions prune hard vetoes; compaction is untrusted rewrite). Re-read those two notes before you treat OpenCode as set-and-forget.

Pitch IDE+platform, not “another OpenCode.” Gateway/account is convenience; BYOK stays. Marketplace and modes are the differentiator vs plain CLI harnesses. Enterprise governance rides Anaconda — evaluate if that middle layer matters for your org.

Vs Claude Code / Cursor / OpenCode / Gemini CLI: this is open-model/BYOK terminal, not a hosted Max seat. 2026 rewrite — Rust Codex fork; classic Python AGPL lives at endolith/open-interpreter community fork. The old pip interpreter UX is no longer current upstream.

Vs Gemini CLI (OSS @google/gemini-cli): Antigravity is the product successor for free/Pro/Ultra consumers — not a 1:1 flag clone (MCP config paths differ). Vs Claude Code / Cursor / OpenCode: Gemini-bound Google stack, not BYOK-agnostic. Enterprise Gemini CLI API/Vertex paths can remain; check which quota you are on.

Vs Claude Code / Codex / OpenHands (coding agents): Spec Kit is the SDD workflow primitive — process and toolkit, not a coding agent seat. You still bring the agent, models, and repo discipline. GitHub org project, not a hosted product SLA.

Vs Semgrep/Gitleaks (deterministic SAST/secrets): OCR adds an LLM agent layer for context-heavy review, not a drop-in replacement for static rules alone. Vs Claude Code/SWE-agent/OpenHands (general coding agents): OCR is a purpose-built review harness, not a full autonomous coder. BYO LLM cost — you bring the model endpoint and pay the provider; the OSS CLI itself is free.

Vs Spec Kit (spec-driven coding workflow), DSPy (prompt/program optimizers), Promptfoo (eval/red-team): OpenResearch is a research-agent harness and experiment ledger around existing coding agents — not a spec kit, not a prompt compiler, not an eval suite. You still pay (or self-host) the coding agent and any remote compute you dispatch.

Vs Worktrunk/jj (git worktree / change-oriented VCS plumbing): Atlas is agent session+change tracking with shared memory, not a general VCS replacement. name collision with Ariga Atlas (slug `atlas`, schema-as-code) — this card is `atlas-agents` / pacifio only; never confuse with atlasgo.io.

Needs an AI coding-agent host — OpenSpec is tooling around the agent, not a product that ships code by itself. Vs Spec Kit (heavier phase-gated SDD): OpenSpec is lighter and more iterate-friendly, still not a standalone SaaS. You still pay (or self-host) the coding agent that runs the slash commands.

Skill only — not a standalone SaaS scanner. Distinct from mcp-security-audit (awesome-copilot skill for .mcp.json / MCP config audits) and from OWASP ASI checklist skills: this is Cloudflare’s multi-phase code/security audit skill for a coding-agent host. You still need the agent runtime and any scanners the skill shells out to.

Vs Spec Kit and OpenSpec: those are process layers for agreeing on specs and changes, while Bend is a language/proof surface — process is not proof. Adopting a new language means a smaller ecosystem and a real learning/toolchain cost before the proof gates pay back.

Vs AgentBox and Octop: Orca is an operator-facing coding-agent ADE with worktrees and design/mobile surfaces, not a generic sandbox/VM launcher or a self-hosted multi-user agent platform. BYO agent CLIs, credentials, model spend, and the operational discipline to reconcile parallel worktrees.

Vs Orca and AgentBox: T3 Code is the remote-control surface for existing agent CLIs, not Orca's worktree/ADE workflow or AgentBox's sandboxed VM launcher. Authentication, local process lifetime, network exposure, and the permissions of the already-signed-in CLI remain your responsibility; a remote UI does not reduce that blast radius.

Engine-linked runtime directories (engine/, proxy/, rewriter/, and related) are Business Source License 1.1, not MIT — see LICENSING.md / LICENSE.BSL before embedding or offering a hosted Caveman runtime. Vs LiteLLM: Caveman is output/style compression for agent hosts, not a multi-provider model gateway. Terse output can hide nuance; keep quality checks on hard tasks.

Vs Pi, OpenCode, and OpenHands: this is the PrimeIntellect product, not the Pi coding agent, a general terminal wrapper, or a full autonomous developer platform. Prime Agent is not a security sandbox — its Python execution runs with the permissions you give it. Isolate it yourself before handing it credentials, production access, or untrusted tasks.

Vs Cua, OpenClaw, and Browser Use: GenericAgent is the tiny self-evolving seed with full system control, not a browser/computer-use protocol, messaging agent, or browser automation framework. Arbitrary OS and ADB control means a bad instruction can reach the host or a connected device. The Chinese README installer also uses plain HTTP — inspect and pin what you execute; do not grant it secrets by default.

Vs T3 Code: Codex-X is Codex-specific visual configuration plus MCP/skills, not a broad remote coding environment or an independent coding model. It handles `auth.json`, supports third-party relays, and allows unrestricted prompt templates — review where credentials go, which endpoint receives requests, and what instructions can execute before using a release. Treat relays as a trust decision, not a convenience toggle.

Vs Excalidraw: this is a human+agent design/decision workspace wired to coding agents, not a general hand-drawn whiteboard. Extreme overnight star fill (~589) is discovery noise; a hosted team product is planned — keep the self-hostable OSS desktop as the free bar until that SKU is real.

Not a generic GGUF runner. Project quants only: DeepSeek V4/V4.1, GLM 5.x, Qwen3.8. Vs Ollama / Magnitude / llmfit: Ollama is the general local model runner, Magnitude is the consumer profiler plus agent-harness wiring, llmfit only sizes models to hardware — DwarfStar 4 is the antirez in-process coding agent with a shared KV cache, not a general local runner. Still beta; pin a commit before a fleet bet.

Model calls still bill Anthropic, Bedrock, Vertex, or Foundry — there is no free model. Vs Claude Code: that gem is the local terminal CLI (paid Max or API); this catch is the GitHub Action that runs on your runner, not the local CLI. Action permissions and secrets are the real gate; pin tag v1.