oh-my-kimi (OMK, Open Multi-Agent Kit) is an open-source terminal coding agent CLI, positioned as "make 'done' pass a single check." Its core idea is: AI saying "done" does not equal acceptance passing.
Purpose and core idea Through /goal verify, the user chooses the command used to check the goal. OMK runs that command after each round, and the goal is only considered complete when it passes in the current workspace. Any later edit invalidates this evidence. The README clearly states that the strength of the result depends only on the chosen check command, and recommends using tests already in the project or a small reproducible demonstration of the goal. This is an explicit workflow; ordinary prompts will not enable this gate.
Switching models within a single session The user chooses a model, works on the repository, and can then use /model to switch to a model better suited to the next step, with the conversation remaining in the same session. It is also possible to stop via /resume or omk -c and return later. The README states that this is manual model switching, not parallel agents or an independent correctness check.
Agents and optional capabilities OMK is a standalone CLI, not a plugin for Claude Code or OpenCode. It supports subscription-based providers, API keys, and local models. By default it starts with one agent that can read files, edit code, and run commands; subagents or explicit verification workflows are added when needed, and are not required for the first task. The README explains that after a fresh install, once provider configuration is complete, only one agent/tool loop starts; it does not turn every prompt into a multi-agent workflow, nor does it automatically certify its answers; the internal lane launcher and automatic command sharding primitives are not wired into the default CLI path. The multi-agent control plane is intended for explicitly configured workflows.
Installation and login A standalone binary can be installed (no Node.js required; the installer verifies the release's SHA256SUMS and refuses if it does not match). With Node.js 22.19 and above, npm is available: npm install -g open-multi-agent-kit --ignore-scripts, or run npx --ignore-scripts open-multi-agent-kit without installing globally. On first run, /login opens, allowing login via subscription or API key; existing Claude Code or Codex CLI logins canReuse via omk provider adopt. Use /model at any time to select other available models.
Security boundaries The built-in local bash uses the operating system sandbox by default: macOS uses sandbox-exec, Linux requires bwrap plus unprivileged user namespaces (on Debian/Ubuntu, sudo apt install bubblewrap). It restricts writes to the workspace and the operating system temporary directory, disables network access, and fails closed with sandbox.backend_missing when the backend is missing. The README emphasizes that this is not read confidentiality or whole-process isolation; other file tools, extension and custom tool code, injected or remote BashOperations, and the OMK process itself still retain the permissions of the process running them, and containerization should be used if the boundary needs to cover more. omk doctor can check the runtime, credentials, the model that the first session will select, and the sandbox, and print remediation methods.
Evidence and benchmark notes The README provides a dated comparison: OMK and mini-SWE-agent on Terminal-Bench 2.1 using the same base model. The observed success rate difference is +4.2 percentage points, 95% confidence interval [−1.9, +10.2] (task-level paired bootstrap), the interval contains zero, so it is not statistically significant; the time budgets are also unequal, with OMK at a median of 482 seconds per round and mini-SWE-agent at 376 seconds. OMK costs about 23% less per round. The README explicitly states that this result does not constitute evidence of overall harness superiority, multi-agent gains, or verification reducing the false completion rate, nor is it an official leaderboard result, and it provides audit corrections, cost definitions, and reproduction instructions.
Repository understanding corpus v0.97.0 introduced the OpenWiki policy and workflow, but did not provide a versioned corpus or integrity verification. v0.98.0 introduced integrity/output guards; the generated corpus is still optional and not packaged. The openwiki/ directory is missing, and the previous untracked corpus was removed because it was proven to contain fabricated evidence (frontmatter symbols defined by 8 undeclared source paths, 45 references to @omk/* package names not published in this repository, and misrepresenting the README's Scope→Route→Verify→Replay loop as a strict engine-likestate machine). CI regenerates the corpus. scripts/check-openwiki.mjs and scripts/check-openwiki-output.mjs are the published guard scripts.
