Grok Buildvskimchi
Generated from the two spec rows. Green marks the better value where a spec has a clear direction. Everything else is just different.
Grok Build
xAI · Terminal agent
#82MCP
Panel
6.30 spec wins
- Reliability
- 6.2
- Usefulness
- 6.8
- Cost
- 5.8
- Longevity
- 6.2
“The old coding model's name now survives only as an alias, which is the software equivalent of a forwarding address.”
kimchi
kimchi · Terminal agent
#75OSSMCP
Panel
6.21 spec wins
- Reliability
- 6.2
- Usefulness
- 6.8
- Cost
- 5.8
- Longevity
- 5.8
“It divides work between an orchestrator, a builder and an explorer, which is more organisational structure than some of its users have.”
Spec by spec
| Spec | Grok Build | kimchi |
|---|---|---|
| Architecture | ||
| Category | Terminal agent | Terminal agent |
| Runs | local, sandbox | local |
| Platforms | macos, linux, windows | macos, linux, windows |
| Context window | 500k tokens on grok-4.6, 256k on grok-build-0.1 | 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 | No | No |
| Sandboxed execution | NoSandboxing uses OS primitives, Landlock on Linux and Seatbelt on macOS, rather than containers; it is off by default and must be enabled, and network blocking is enforced on Linux only. | No |
| Multi-agent orchestration | YesBuilt-in general-purpose, explore and plan subagents run as independent child sessions. | YesMulti-model mode assigns separate models to orchestrator, builder and explorer roles and delegates tasks between them. |
| Headless / CI mode | YesThrough the -p flag with streaming JSON output. | YesThe repository ships a Terminal-Bench adapter that aggregates usage across session files, which implies unattended runs. |
| Models | ||
| Backbone | Grok | kimchi-dev models, Anthropic |
| Bring your own model | YesCustom model entries can be added in ~/.grok/config.toml, but a self-hosted or local base URL is not documented. | Yes |
| Local models | No | No |
| Cost | ||
| Pricing model | usage | usage |
| Starts at | n/a | n/a |
| Free tier | No | No |
| Bring your own key | Yes | Yes |
| Openness | ||
| Open source | No | Yes |
| License | proprietary | Apache-2.0 |
| GitHub stars | n/a | 2,231 |
Which one would each critic pick
| Critic | Grok Build | kimchi | Pick |
|---|---|---|---|
| El Juez | — | — | not enough reviews |
| El Amigo | 6.3 | 6.5 | kimchi — Pick it if you want role-splitting without wiring it yourself; pick a single-model agent if you would rather not debug three models at once. |
| El Crítico | 6.3 | 6.0 | Grok Build — Sandboxing is off until you turn it on, and the network half of it is enforced on Linux only, so a macOS session has weaker containment than the docs suggest at a glance. |
| El Profesor | 7.0 | 6.5 | Grok Build — Two structural choices carry the design: subagents run as independent child sessions with their own context, and the agent embeds in other editors over the Agent Client Protocol. |
| La Inversora | 6.8 | 6.0 | Grok Build — This is a model company's distribution channel wearing an agent's clothes; the product's job is to make its own tokens the default purchase. |
| La Jefa | 5.8 | 5.0 | Grok Build — There is no seat price at all: at two dollars per million in and six out, sixty engineers is a meter with no ceiling and no per-person cap. |
| El Hacker | 5.5 | 7.0 | kimchi — Apache-2.0, MCP tools attach, and an external provider can replace the built-in models, so the vendor's endpoint is a default rather than a cage. |
Picks are derived from each critic's own scores. Humans vote on matchups on the duels page.