karotte

karotte runs LLM agents on tasks and scores the results.

You write an environment: a Python project with one or more tasks. A task is a list of steps. Each step gives the agent instructions and a judge that decides whether the agent succeeded. karotte builds the environment into a container image, lets the model work inside it through tools like bash, and records every message, tool call and score in a transcript.

The agent runs as an unprivileged user with its own resource limits, a firewall, and a disk quota, so a task can hand it a real shell without trusting it. The model under test is called the student.

Install

karotte needs Python 3.12+, uv, and docker with buildx (or podman with --runtime podman) for containerized runs. Ubuntu's docker.io lacks buildx; install docker-buildx too. The justfiles need just 1.40 or newer; uv sync --extra dev installs one into the venv.

uv tool install karotte

Quick start

Create an environment from the default template and run its example task:

karotte create-env my_env
cd my_env
uv sync --extra dev
uv run setup_data.py
uv run karotte create-run-config --model claude-fable-5
export ANTHROPIC_API_KEY=...
uv run karotte run --config run_config.json

Commands that load your tasks (create-run-config, run, tasks list, check) have to run in the environment's venv, hence uv run. setup_data.py prepares data the image needs, such as model weights; the default template's version does nothing. karotte check loads every task and tool and fails if one doesn't load; the image build runs it too.

karotte run builds the image, runs the task, and writes the transcript to out/transcript.json. Files the task saves as artifacts go to out/<run_id>_artifacts/. karotte dashboard out/ shows the transcripts in out/. karotte models list shows which models karotte knows; model ids are passed to litellm. The API key is model_api_key in the run config. create-run-config sets it to the key variable of the model's provider, such as $ANTHROPIC_API_KEY or $OPENAI_API_KEY, which karotte reads from your environment when the run starts. --model-api-key sets another.

Useful while you work on a task:

• uv run karotte run --config run_config.json --dev skips the image build and mounts your src/environment/ into the container.
• uv run karotte tasks list shows the tasks the environment defines; create-run-config --task <task-id> picks one.

The container image

Every karotte run without --dev rebuilds the image first. podman and docker reuse cached layers, so an unchanged environment builds quickly.

--dev skips the build and runs the existing image, with your src/environment/ mounted over the installed copy. Rebuild after changing dependencies, the Containerfile or data.

karotte build builds the same image without running a task. run always uses the image tagged karotte, which is build's default tag, and has no option to pick another. build --tag is for images you use elsewhere.

Model endpoints and proxies

karotte run --proxy <url> sends the model calls to <url> instead of the provider. Pass the root URL, without /v1:

• Claude models go to <url>/v1/messages. karotte sets ANTHROPIC_BASE_URL to <url> in the container, and litellm adds /v1/messages.
• Other models go to <url>/v1, so the endpoint needs an OpenAI-compatible API there. vertex_ai/ models ignore the proxy.

The same flag points karotte at any endpoint that serves these paths. The URL is used inside the container, where localhost is the container itself. Variables like ANTHROPIC_BASE_URL in your shell don't reach the container. With --no-containerized, karotte sets ANTHROPIC_BASE_URL and KAROTTE_PROXY_URL in its own process instead, unless one of them is already set. --no-containerized only runs inside a karotte container image, where KAROTTE_CONTAINERIZED is set.

A key is needed only if the endpoint asks for one. karotte still sends model_api_key; if that names a variable such as $OPENAI_API_KEY that is unset, --proxy fills in a placeholder.

Without --proxy, karotte uses the URL a plugin registers under karotte.default_proxy_url (see Plugins). Without such a plugin, calls go straight to the provider. --no-proxy ignores the plugin's URL.

KAROTTE_INFERENCE_SERVICE_TIER=priority asks Fireworks and Vertex AI Gemini for their priority tier, auto only once the provider runs out of capacity. Unset, the default, uses the provider's default tier.

Writing tasks

Each task is a package in src/environment/tasks/: a directory whose __init__.py defines the Task class. Plain files and directories starting with _ are skipped. uv run just create-task <task-id> copies the task template (just comes with uv sync --extra dev and lives in the venv). A minimal task:

import re

from karotte import Step, Task
from karotte.judges import RegexJudge


class FindPython(Task):
    id = "find-python"

    @property
    def system_prompt(self) -> str:
        return "You are working in a Linux shell."

    @property
    def steps(self):
        return [FindPythonStep(config=self.config)]

    @property
    def tools(self):
        return ["bash"]


class FindPythonStep(Step):
    @property
    def instructions(self) -> str:
        return "Find the path to your Python executable. Answer with `path: <path>`."

    @property
    def judge(self):
        return RegexJudge([re.compile(r"path: .*/python3?")])

Judges in karotte.judges: RegexJudge matches the final message, ExecutableJudge runs a scoring script, RubricJudge asks an LLM to grade against a rubric, and Judge is the base class for your own. The example task in the default template shows submissions, hints, and hooks that run before scoring.

Templates

A template is a directory of files, rendered with Jinja into a new environment. karotte ships two:

• default: a CPU environment with an example task. Every other template builds on it.
• language-toolchains: gives the agent exactly one language toolchain per task.

karotte templates list shows every installed template. Templates stack: karotte create-env my_env --template default --template language-toolchains renders default and then language-toolchains on top. A later template can replace a file or override a Jinja block in it:

{% extends "default/CLAUDE.md" %}
{% block claude_md -%}
New content
{% endblock %}

karotte update brings an existing environment up to the latest templates with a 3-way merge, so your own edits survive.

Writing a template package

Templates can live in their own Python package. Put the template directories (each with a template.toml) under one directory and register it under the karotte.templates entry point:

[project.entry-points."karotte.templates"]
my_templates = "my_package:TEMPLATES_DIR"

TEMPLATES_DIR is a path to that directory. A template.toml holds the fields of EnvironmentTemplate, for example:

description = "Adds a Rust toolchain."
requires = ["default"]

Install the package next to karotte (uvx --with my-package karotte create-env ...). karotte records it in the environment's manifest, so karotte update pulls it in again.

Plugins

Besides templates, a package can extend karotte through these entry points:

A plugin that fails to load is skipped with a warning, except a run config preprocessor: that one fails the run.

Connecting a backend

Without a backend, karotte writes the transcript to a file. To collect runs centrally, set backend_uri in the run config. karotte then calls these HTTP endpoints on it, with run_id as a query parameter (presign has it in the body):

With the external agent the backend also supplies the model's messages: GET /api/internal/get_message returns the next message (404 while there is none) and POST /api/internal/delete_message consumes it.

Requests carry Authorization: Bearer <token> if the file at KAROTTE_BACKEND_TOKEN_PATH (default /var/run/secrets/service-account-token) exists. The events are the models in karotte.schemas.transcript.

Development

just lint
just test
just test-template default

Every merge to main is released. The patch version is the commit count.

License

karotte is under the MIT license. The templates in src/karotte/templates/ are under MIT No Attribution (MIT-0), so environments created from them need no license notice.

Third-party software in built images

Images built from the templates contain third-party software under its own licenses. They are based on Amazon Linux 2023. The language-toolchains template installs GPL and LGPL software (gcc, GnuCOBOL, GNU Prolog, Free Pascal, the libraries bundled with Julia, and others), Amazon Corretto (GPL-2.0 with the Classpath Exception) and Clojure (EPL-1.0) for the languages you enable. Mojo, off by default, is under Modular's proprietary license and has its own terms even if you never share the image. The rest only matters if you redistribute a built image, for example by pushing it to a public registry.