Skip to content

Latest commit

 

History

History
79 lines (51 loc) · 7.83 KB

File metadata and controls

79 lines (51 loc) · 7.83 KB

code-oz vs other AI coding tools

This page exists because every developer who finds code-oz immediately asks: "why not just use Claude Code, Codex, Cursor, or Aider directly?" The answer is that those tools are coding agents; code-oz is a governed delivery loop around them. The two are complementary, not competitive.

The table below compares mechanics, not marketing. Every row corresponds to a feature you can verify in the linked competitor's own documentation. Every footnote cites its source. The table was authored by code-oz maintainers and independently reviewed by an external gpt-5.5 reviewer to catch overclaiming; partial-credit ratings (partial) are used wherever a competitor delivers a related capability with different mechanics, so we never claim a competitor lacks a feature they actually ship.

This is the canonical public comparison. The README's short table is a summary that links here.

Feature-by-feature comparison

Feature Cursor Claude Code Aider Continue Devin code-oz
Same SHA-pinned native release across npm/Homebrew/curl partial¹ partial² partial⁸ partial⁹ partial¹³
Orchestrated cross-provider phase roles (different LLM per phase) partial³ partial¹⁰
Cross-family adversarial REVIEW (different LLM family)
File-based gates with sha256 binding partial⁴
Worktree-per-run isolation partial¹¹ partial¹² partial⁴
Full SDLC artifact trail (SPEC/PLAN/BUILD_REPORT/VERIFY/REVIEW) partial⁵ partial⁴
Debate-policy scheduler on disagreement
Run-level cost budget enforcement with kill-switch partial⁶
Runs on CLI auth (no API keys required) partial¹⁴ partial partial¹⁵ partial⁷
Open source (MIT)

Footnotes

¹ Cursor's CLI (cursor-agent) installs via a shell installer per the current Cursor docs (cursor.com/docs/cli/installation), but ships separately from the Cursor IDE; the row stays partial because Cursor's official install surface is a single channel today, not a single SHA-pinned native release across npm/Homebrew/curl.

² Claude Code ships Native Install, Homebrew, WinGet, and Linux package-manager installers per code.claude.com/docs/en/quickstart; each channel is its own installer, not a single SHA-pinned release asset shared across npm/Homebrew/curl.

³ Cursor's published model documentation lists OpenAI, Claude, and Gemini models (cursor.com/docs), but Cursor does not assign different providers to different SDLC phases as orchestrated phase roles.

⁴ Devin exposes sessions, PR links, Session Insights/timeline, and audit logs per docs.devin.ai/get-started/first-run + docs.devin.ai/enterprise/api-reference/audit-logs, but does not document file-based gate machinery or worktree-per-run isolation; "partial" reflects an opaque audit trail, not equivalent mechanics.

⁵ Aider captures commit messages + diff history per aider.chat/docs/git.html; not the full gated artifact set that code-oz records.

⁶ Cursor documents account- and plan-level usage limits in its usage documentation (cursor.com/docs), not per-orchestration-phase budgets.

code-oz's Claude and Codex providers run through their upstream CLI login sessions, so they need no API key; xAI requires XAI_API_KEY. The row is partial because not every provider is keyless.

⁸ Aider distributes via shell installer + uv + pipx + pip per aider.chat/docs/install.html — Python packaging channels, not native binaries pinned across npm/Homebrew/curl by SHA.

⁹ Continue CLI installs via shell installer and npm per docs.continue.dev/cli/quickstart; the install row is partial because the same binary is not released as a SHA-pinned asset across Homebrew, curl, and npm together.

¹⁰ Aider's Architect mode pairs a separate architect model with an editor model per aider.chat/docs/usage/architect.html; that is a two-model split inside one edit, not different providers orchestrated across distinct SDLC phases, so the row is partial.

¹¹ Cursor documents isolated worktrees for background agents per cursor.com/cli and the 3.2 changelog (cursor.com/changelog/3-2); worktree isolation exists, but it is not bound to a gated run lifecycle, so partial.

¹² Claude Code documents claude --worktree, automatic desktop worktrees, and subagent worktree isolation per code.claude.com/docs/en/setup; worktree isolation exists, but it is not bound to a per-run gate lifecycle, so partial.

¹³ Devin for Terminal ships CLI installers per docs.devin.ai/get-started/first-run and cli.devin.ai; that is an install channel, not the same SHA-pinned native release shared across npm/Homebrew/curl, so partial.

¹⁴ Cursor's agent CLI documents cursor-agent login browser authentication alongside an API-key option per cursor.com/docs/cli/reference/authentication; CLI-session auth exists, so partial rather than ❌.

¹⁵ Devin documents devin auth login per cli.devin.ai/docs/enterprise/devin-auth, though the flow is enterprise-scoped; CLI-session auth exists, so partial rather than ❌.

Best used with

If you already use... code-oz adds
Claude Code The same Claude as your builder, plus a Codex (or other family) reviewer at the REVIEW gate; SHA-bound approvals; an inspectable lifecycle trail.
Codex CLI Same pattern with Codex as builder; Claude (or other family) as the cross-family reviewer.
Cursor An external lifecycle layer outside the editor — useful when a change needs an audit trail beyond what the IDE captures.
Aider A multi-phase artifact set on top of Aider's git-native model edits, with cross-family review and approvals.
Devin / Factory Local-first lifecycle governance for changes you want to keep on your machine.
Qodo / Sonar An EARLIER gate stage — the SPEC, PLAN, and approval artifacts produce inputs that Qodo / Sonar then review at PR time.

What code-oz is NOT

  • Not a coding agent itself. The persona prompts orchestrate Claude / Codex / xAI / Fake. They do not contain a model.
  • Not a replacement for direct agent use. Use Claude Code or Codex directly when you want the fastest possible loop. Use code-oz when you want a governed loop around them.
  • Not a static analyzer. Sonar, Qodo, ESLint, and Bandit do that better.
  • Not an enterprise SaaS. No hosted runtime, no team dashboard, no RBAC. Local CLI only.
  • Not a benchmark for "which model writes better code." The deterministic FakeProvider demo proves lifecycle gates and ledger determinism, not model quality. The Agent Gate Bench protocol is the path to measured comparisons; the runner shipped in v0.21 with the first measured code-oz Fake baseline rows. The direct-agent and live-provider columns require local credentials and are not yet measured.

Methodology note

Footnote sources were verified by an external gpt-5.5 review against the competitor's own published docs, last re-verified 2026-05-21. The 2026-05-21 pass softened several competitor ratings to partial after confirming current docs: Cursor and Claude Code worktree isolation, Cursor CLI auth, Devin install and CLI auth, and the Aider architect/editor model split. If a competitor ships a feature that this table marks as or partial, that is a documentation gap on our side — please open an issue with a link to the competitor's published doc.

For the underlying provider matrix code-oz ships today (Claude / Codex / xAI / Fake live; Gemini stub; OpenCode and Roo Code as future candidates), see docs/contracts/PROVIDERS.md § "Provider status (v0.1)".