A coding agent in Go. Two binaries: kvit-coder runs one turn (user question -> multiple steps calling tools -> model's answer) and exits. And kvit-coder-ui - the interactive terminal front end.
The agent started as experiment to see how much coding can be done with locally ran weak models and was extended with features more suitable to strong models. Weak/strong modes change how tools behave and how much handholding is done. Each model added to the agent flagged as weak or strong.
It uses small amound of RAM (<30 Mb) compared to hundreds Mbs other agents use. See docs/agents-ram.md.
This is important to me as I often have dosens of agents opened/running/waiting at any moment.
curl -fsSL https://raw.githubusercontent.com/kvit-s/kvit-coder/main/install.sh | shTo build from source, see CONTRIBUTING.md.
Windows (native, 10/11). Download the installer for your machine
(kvit-coder_<version>_windows_amd64_setup.exe, or arm64) from the
latest release and run
it. It installs for your user only, without an administrator prompt, into
%LOCALAPPDATA%\Programs\kvit-coder, adds that folder to your user PATH, puts
kc.exe and kcu.exe beside kvit-coder.exe and kvit-coder-ui.exe as the
short names, and seeds %USERPROFILE%\.kvit-coder\config.yaml when there is
none. Open a new terminal afterwards so it sees the new PATH. Uninstall it from
Settings > Apps; that removes the PATH entry and leaves
%USERPROFILE%\.kvit-coder alone.
Without the installer, download the portable zip
(kvit-coder_<version>_windows_amd64_portable.zip, or arm64), extract it
where you want to keep it (e.g. C:\apps\kvit-coder), then run the
install.ps1 inside that folder. It does the same for the folder it runs from:
powershell -ExecutionPolicy Bypass -File install.ps1Or install a published release straight into %USERPROFILE%\bin, which the
script also puts on the PATH:
irm https://raw.githubusercontent.com/kvit-s/kvit-coder/main/install.ps1 | iexPlatforms. Linux is where this is developed and used daily, including under
Windows Subsystem for Linux (WSL). The macOS builds compile and are published,
but never tested. Native Windows (10/11, amd64 and arm64) builds, runs, and passes tests;
it needs Git for Windows (for sh.exe, grep, and git itself) and installs
via the installer, the portable zip or install.ps1 (see above). WSL stays supported regardless; benchmarks stay
WSL/unix-only (see spec/windows.md).
Run kcu. With no model configured it opens :setup: choose OpenCode Zen,
OpenCode Go, a local llama.cpp, vLLM, Ollama or LM Studio server, or any other
endpoint that speaks the same protocols; paste its key; and tick the models
you want. It fills in each model's protocol, context size and effort levels
from the endpoint and the models.dev catalog, saves them
to ~/.kvit-coder/models.yaml and the key to ~/.kvit-coder/credentials.json,
and leaves config.yaml alone. Once a model is saved it offers web search,
which needs a Brave Search API key. :setup adds more later and turns web
search on or off, :models switches between models, and :keys replaces a
key; see docs/cli.md.
The installers put a copy of config.example.yaml at
~/.kvit-coder/config.yaml (on Windows: %USERPROFILE%\.kvit-coder\config.yaml),
and kcu offers to write it when there is none. It holds everything except the
model; a model can also be written into it by hand
(docs/models.md).
Without -config, both binaries take the first of $KVIT_CODER_CONFIG,
./config.yaml, ~/.kvit-coder/config.yaml (%USERPROFILE%\.kvit-coder\config.yaml
on Windows), and config.yaml beside the binary with symlinks resolved.
Every key is documented in
docs/configuration.md, including the parts
config.example.yaml leaves out: defining more than one model and switching
between them, Model Context Protocol (MCP) servers that provide external tools,
tool groups that hide tools until the model asks for them,
and the path and shell safety rules.
# Interactive. The directory it starts in is the workspace.
kcu
# One turn, headless. This is what you script against.
kvit-coder -p "Find all TODO comments in the codebase"
# Only the final answer on stdout.
kvit-coder -pq "What does main.go do?"A turn that fails exits non-zero and says why. On a terminal the final answer is
rendered as styled markdown; piped output, --json, NO_COLOR and TERM=dumb
stay raw. Every flag is in docs/cli.md.
This program executes shell commands and edits files. workspace.path_safety_mode
decides what happens when a path outside the workspace directory is read or
written: allow, block, warn, ask_once or ask_always, where the asking
modes prompt on /dev/tty and remember the answer for the run. Shell commands
are parsed into a syntax tree and each simple command in the line is checked on
its own against the built-in rules and the allow and deny lists in
tools.shell. A refused command comes back to the model as a tool result saying
what it would have done, so the turn can continue with something else. --yolo
disables these prompts for a run.
The harness includes three benchmarks. These are designed to check basic agentic
skills of weak models. See benchmarks/README.md.
docs/ holds the reference documentation: the
tools the model can call, every
command-line flag, every configuration key,
how to add an OpenCode model,
sessions and steering, external tools, the
memory measurements quoted above, and
how the program is built: one process per turn, sessions
as directories, and the other design decisions behind it.
MIT