CrushvsWrongStack
Generated from the two spec rows. Green marks the better value where a spec has a clear direction. Everything else is just different.
Crush
Charm · Terminal agent
#25OSSMCP
Panel
6.51 spec wins
- Reliability
- 6.0
- Usefulness
- 6.5
- Cost
- 7.3
- Longevity
- 6.2
“Has a --yolo flag, auto-discovers your local llama.cpp, and a license with MIT in the name that is not MIT yet.”
WrongStack
WrongStack · Terminal agent
#85OSSMCP
Panel
6.42 spec wins
- Reliability
- 5.7
- Usefulness
- 6.8
- Cost
- 7.7
- Longevity
- 5.3
“It ships six different user interfaces, so there is now a wrong way to run it to suit every mood.”
Spec by spec
| Spec | Crush | WrongStack |
|---|---|---|
| Architecture | ||
| Category | Terminal agent | Terminal agent |
| Runs | local | local |
| Platforms | macos, linux, windows | macos, linux, windows |
| Context window | not documented | not documented |
| Protocols | ||
| MCP client | Yes | Yes |
| MCP server | No | No |
| Capabilities | ||
| Runs terminal commands | Yes | Yes |
| Multi-file edits | Yes | Yes |
| Git operations | Yes | Yes |
| Browser control | NoThe built-in tools stop at web_fetch, web_search and download; there is no browser, CDP or Playwright tool. | YesBrowser and end-to-end testing are among the 61 first-party built-in tools listed in the comparison table. |
| Sandboxed execution | NoShell and edit calls are guarded by permission prompts (or skipped with --yolo), with no container or OS-level isolation. | No |
| Multi-agent orchestration | No | Yes |
| Headless / CI mode | Yes`crush run` executes a single non-interactive prompt, reads stdin and supports --quiet for scripts. | YesA single-shot non-interactive query mode is documented; there is no dedicated CI action. |
| Models | ||
| Backbone | Claude, GPT, Gemini, any OpenAI-compatible or Anthropic-compatible provider | ~140 providers via models.dev, Anthropic, OpenAI, GitHub Copilot, Ollama, vLLM, LM Studio |
| Bring your own model | Yes | Yes |
| Local models | YesOllama, LM Studio, llama.cpp and MLX providers are auto-discovered, and any OpenAI- or Anthropic-compatible base_url can be added. | YesOne-command presets exist for Ollama, vLLM and LM Studio, and any custom base URL can be used to run fully on localhost. |
| Cost | ||
| Pricing model | byok | byok |
| Starts at | $0/mo | $0/mo |
| Free tier | Yes | Yes |
| Bring your own key | Yes | Yes |
| Openness | ||
| Open source | Yes | Yes |
| License | FSL-1.1-MIT | MIT |
| GitHub stars | 28,444 | 356 |
Which one would each critic pick
| Critic | Crush | WrongStack | Pick |
|---|---|---|---|
| El Juez | — | — | not enough reviews |
| El Amigo | 7.3 | 6.3 | Crush — Crush is the terminal agent to pick if you want LSP-grade context and a Charm-quality interface with your own keys, and the one to read the license on before you fork it. |
| El Crítico | 6.3 | 5.3 | Crush — A --yolo flag on an agent with no sandbox means a headless run is your shell with nobody watching; the LSP context is the reason to read past it. |
| El Profesor | 6.5 | 6.8 | WrongStack — Tasks are gated on atomic verification before advancing, and a policy layer classifies each risky choice as automatic, denied or escalated, which are two different control mechanisms. |
| La Inversora | 5.8 | 5.8 | no preference |
| La Jefa | 5.8 | 6.0 | WrongStack — Nothing per seat, and a command centre aggregates sessions, agents, spend and worktrees across machines with per-agent budgets, which is the closest thing to a console anyone here ships. |
| El Hacker | 7.5 | 8.3 | WrongStack — MIT, roughly 140 providers pulled live from a public catalogue, presets for three local serving stacks, and a flag that boots the kernel fully offline with the extras switched off. |
Picks are derived from each critic's own scores. Humans vote on matchups on the duels page.