Kode CLI ships five bin entries, skips permissions by default, and points at an empty releases page
Kode CLI — Design for post-human workflows. One unit agent for every human & computer task.
At a glance
- What is it?
- Kode CLI is a terminal coding agent published to npm as @shareai-lab/kode. Its distribution, permission defaults, instruction discovery and network paths each come with a cost the documentation states plainly.
- Who is it for?
- Take Kode CLI if you want one install that reads AGENTS.md, edits files and runs shell commands, and if you can live with starting on --safe inside a tree you trust. Leave it if your build strips npm optionalDependencies and leaves no system rg, or if you need the portable single-file binary while the releases page stays empty.
- Can I use it commercially?
- Yes. Apache-2.0 is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 13 days ago.
- What is it written in?
- Mainly TypeScript, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.
Star history
- shareAI-lab/Kode-CLI5,236 ★ · +8
| Repository | 2026-09-23 | 2026-11-15 |
|---|---|---|
| shareAI-lab/Kode-CLI | 5.2k | 5.2k |
8 Stars in the last 53 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
Three command names share one script, two more bins ship with no prose behind them
package.json registers five bin entries, and the README's post-install command list names three of them. `kode`, `kwa` and `kd` all resolve to `cli.js`, so the three names are one program with different letters on the command line, not three modes. The other two are `mcp-cli`, which points at `mcp-cli.js`, and `kode-acp`, which points at `cli-acp.js`.
That matters when you unpack the published tarball looking for a way to drive Kode from another process. Judging by their names, two of the five entry points are MCP and ACP surfaces. The package does export entry points for embedding, `./protocol` resolving to `dist/sdk/protocol.js` and `./daemon-client` to `dist/sdk/daemon-client.js`, each with a CommonJS twin. `examples/daemon-client-echo.ts` is the one example file that looks like a runnable client, and `examples/agents/` is the other example entry.
YOLO mode is the default, so the first prompt already holds write access
The default permission posture is the opposite of cautious, and the security notice in the README says so without hedging: Kode runs in YOLO mode by default, equivalent to the `--dangerously-skip-permissions` flag, bypassing all permission checks. The way back is a single flag.
kode --safeWith `--safe`, permission checks and manual approval apply to all operations. The notice scopes YOLO to trusted, secure environments and non-critical projects, and points at `--safe` when the files matter or when the model's capability is in doubt.
What YOLO cannot do is ask. Every check is bypassed rather than deferred, so nothing stands between a model decision and your working tree, and the cost of a badly aimed first prompt lands on files instead of in a dialog. Treat the flag as part of the install rather than a later refinement: on a repository with uncommitted work, run `--safe` until you have watched what the tool does with the tree.
The releases page the README points to has no assets on it
The portable path is described as an option: download the Bun-compiled single-file asset from GitHub Releases, with `docs/binary-distribution.md` holding asset names and local build steps. The repository has no GitHub releases published, so that download has nothing to land on today.
The build recipe is real even though the download is not. The Dockerfile starts from `node:22.14.0-alpine`, installs `[email protected]` globally, and runs a four step build with a frozen lockfile:
bun install --frozen-lockfile
node scripts/build-seccomp-assets.mjs --require
bun run build
npm pack --ignore-scripts --pack-destination /tmpThe runtime stage adds `ripgrep` through apk, installs the packed tarball with `--omit=optional --ignore-scripts`, sets `WORKDIR /workspace`, drops to `USER node`, and uses `kode` as its entrypoint.
Plan around two limits. The image deliberately takes the `--omit=optional` path, so it runs on the Node.js entry `dist/index.js` and depends on the system `rg` it installed itself. Locally, `bun.lock` and `packageManager: [email protected]` mean a build outside the image wants that same Bun version, or dependency resolution differs from what the image produced.
optionalDependencies decide whether search runs natively or through Node
Installation is one global npm command, and the interesting decisions happen at the edges of that install:
npm install -g @shareai-lab/kode
kodeTwo per-platform packages arrive through `optionalDependencies`. `@shareai-lab/kode-ripgrep-<platform>-<arch>` supplies ripgrep, and `@shareai-lab/kode-bin-<platform>-<arch>` supplies a native CLI binary. Kode prefers the native binary package and falls back to Node.js when it is missing, and nothing is downloaded from GitHub during install because the standalone binaries live on the releases page, separate from npm.
That optionality is also the failure mode. With `--no-optional` or `--omit=optional` the CLI runs through the Node.js entry `dist/index.js`, and the ripgrep package disappears with it, so you need a system `rg` or an explicit `KODE_RIPGREP_PATH` before search works. Locked-down images that strip optionals get the slower path by default, and `engines` requires Node `>=20.18.1`. A separate dev channel exists as `npm install -g @shareai-lab/kode@dev`, and users behind restricted networks are pointed at `--registry=https://registry.npmmirror.com`.
When the ripgrep package is absent, the README's remedy is to install a system `rg` or set `KODE_RIPGREP_PATH`, stated inside an install note.
Instruction discovery concatenates root to leaf under a 32 KiB ceiling
Kode implements the AGENTS.md format and spells out the walk. Instructions are collected by walking from the Git repo root to the current working directory. In each directory it prefers `AGENTS.override.md` over `AGENTS.md`, at most one file per directory, and discovered files are concatenated root to leaf. Combined size is capped at 32 KiB by default, with `KODE_PROJECT_DOC_MAX_BYTES` as the override. A `CLAUDE.md` in the current directory is read too, as a legacy instruction file, and legacy `.claude` directories are supported for some features.
Two consequences follow from that ordering. Because concatenation runs root to leaf under one shared ceiling, a large root file eats into the budget left for the directory you actually work in. And because `AGENTS.override.md` outranks `AGENTS.md` with no merge step, a local override silently replaces the version that was committed, which matters when a teammate's behavior depends on the checked-in text.
The same mechanism backs documentation authoring: `# Your documentation request` generates and maintains an `AGENTS.md` while keeping existing `.claude` workflows readable. The README puts the AGENTS.md standard at 60k plus open source projects. Treat 32 KiB as a budget you spend deliberately, and list the override files present in a repo before concluding the model ignored instructions it never received.
No telemetry by default still leaves four networked paths open
The privacy note is narrow and worth reading precisely. Kode sends no product telemetry or analytics by default. Requests happen when you use networked features: model provider calls against the Anthropic or OpenAI-compatible endpoints you configure, the `WebFetch` and `WebSearch` tools, plugin marketplace downloads from GitHub or URL sources along with OAuth flows, and optional update checks you opt into with `autoUpdaterStatus: enabled`.
So the default-off claim covers analytics, not model traffic or plugin downloads. The repository's `.env.example`, used for production API tests, shows what test endpoint settings look like:
PRODUCTION_TEST_MODE=true
TEST_GPT5_API_KEY=your_gpt5_api_key_here
TEST_GPT5_BASE_URL=http://127.0.0.1:3000/openai
TEST_MINIMAX_API_KEY=your_minimax_api_key_here
TEST_MINIMAX_BASE_URL=https://api.minimaxi.com/v1A base URL of `http://127.0.0.1:3000/openai` is a local endpoint in the OpenAI-compatible shape, which is also how a self-hosted model server would be reached. Global state, including models, pointers and theme, lives in `~/.kode.json`, or `<KODE_CONFIG_DIR>/config.json` when that variable is set, while project settings go to `./.kode/settings.json` and `./.kode/settings.local.json`.
What Kode cannot do is block those calls for you. The guarantee you get is the absence of analytics; everything else on the network depends on which features you touch and which base URLs you type in.
Model selection decides the outcome before any flag does
Before tuning anything else, the README offers a blunt model rule: use newer models built for autonomous task completion, and avoid older question answering models such as GPT-4o or Gemini 2.5 Pro, because they are optimized for answering questions rather than sustained independent task execution. It also asks for models trained on agentic workflows and extended reasoning.
The rest of the model surface is per call rather than per project. `@ask-model-name` consults a specific model for specialized analysis, `@run-agent-name` delegates to a subagent, and `Option+M` cycles the active model. Models are configured through `/model` in the UI or through `kode models import/export` in YAML, with further detail in `docs/develop/configuration.md`.
Prompt authoring has its own bindings. `Option+G` opens the current message in `$EDITOR` or `$VISUAL`, falls back to code, nano, vim or notepad, and returns the text when you close it. `Option+Enter` inserts a newline without sending, and plain Enter submits. Completion covers slash commands, file paths, agents, configured models and available shell commands, with fuzzy and abbreviation matching, and accepting an agent or model entry adds the `@` prefix for you.
What the model layer cannot fix is a chat tuned endpoint. If the configured model stops halfway through a task, changing permission flags will not change the outcome, because the failure sits in the completion style rather than in the tool loop.
Editorial conclusion
Take Kode CLI if you want one install that reads AGENTS.md, edits files and runs shell commands, and if you can live with starting on --safe inside a tree you trust. Leave it if your build strips npm optionalDependencies and leaves no system rg, or if you need the portable single-file binary while the releases page stays empty. Before either, confirm Node is 20.18.1 or newer, check that no AGENTS.override.md is shadowing the instructions you meant to ship, and pick a model trained for agentic work, because permission flags cannot repair a model that stops mid-task.
Frequently asked questions
What is Kode CLI?
Kode CLI is a terminal coding assistant from shareAI-lab, published to npm as @shareai-lab/kode at version 2.2.0. It reads a codebase, edits files, runs shell commands, and delegates work to subagents through @run-agent-name. Node.js 20.18.1 or newer is required.
What does the name Kode mean in Kode CLI?
The repository description states the intent: one unit agent for every human and computer task, designed for post-human workflows, with AGENTS.md as its instruction format and Apache-2.0 as its license.
What does CLI stand for in Kode's packaging?
CLI is the command line interface, and package.json shows what that means here: kode, kwa and kd all map to cli.js, two further bins are mcp-cli and kode-acp, and ./protocol plus ./daemon-client are exported from dist/sdk for embedding.
Which CLI is best for coding with Kode?
The README says model choice decides most outcomes: pick models trained for agentic task completion and skip older question answering models such as GPT-4o or Gemini 2.5 Pro.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/shareai-lab-kode-cli)Star history chart for your README
[](https://hysenlabs.com/projects/shareai-lab-kode-cli)Was this useful?