Free, open tools for AI safety
Sandbox AI agents, watch what they do with a hidden Observer, and measure it with evals. Open source, local, with evidence anyone can check.
Embeddable Python LLM agent core with tools, Agent Skills, MCP, and media generation.
Developer toolsPython · MDXGitHub @cliver-projectOnlineOwner unconfirmed
Have you tried it? One tap is enough. No sign-in needed.
No updates detected yet. GitHub releases will appear here.
The maker's introduction consists of the line above. Read the repository README for more details. Repository README ↗
Loading the full README.
Read the full README ↗CLIver is an embeddable agent core with a companion CLI, built from two packages:
cliver-core — the embeddable engine: an agent loop with tool calling, provider engines
for the OpenAI and Anthropic protocols, per-purpose model bindings (chat, image, video, TTS,
ASR, embed, rerank), nine built-in tools, Agent Skills, MCP, and opt-in permission and
execution-policy axes. No CLI, database, or config files — Python objects in, messages out.cliver — the CLI application built on cliver-core: config-path and skill-discovery
conventions plus a cliver command that runs one task per invocation or starts a streaming
interactive REPL, with interactive tool approval in both modes. It remains experimental and
is not on PyPI yet.cliver-core 0.3 is the first stable series: the public API stays backward compatible, and
0.3.x patch releases contain bug fixes and backward-compatible additions. The cliver CLI is
for tinkering:
run one-shot tasks or a multi-turn REPL with your own skills and tools, approve mutating calls
interactively, and inspect tool events. See
Stability & Versioning
for the support policy.
Install the core library (Python 3.10–3.13):
pip install cliver-core==0.3.1
import asyncio
from cliver_core import AgentCore, ModelBindings, presets
async def main():
bindings = ModelBindings(llm=presets.deepseek(api_key="..."))
async with AgentCore.open(bindings) as agent:
response = await agent.chat("Summarize the agent loop in one sentence.")
print(response.message.content)
asyncio.run(main())
See the cliver-core README for tools, skills, MCP, media capabilities, and the full API.
The cliver CLI lives in this repository (not on PyPI yet). With a prompt it runs one task;
with no prompt on a terminal it starts a streaming REPL:
export CLIVER_MODEL=deepseek-v4-flash
export CLIVER_PROVIDER=deepseek # or CLIVER_BASE_URL for any compatible endpoint
export DEEPSEEK_API_KEY=... # or CLIVER_API_KEY
uv run cliver --yes "Read pyproject.toml and summarize the dependencies."
uv run cliver --yes # interactive REPL: /help, /clear, /exit
--yes approves ASK tools without prompting; without it, safe tools still run and each ASK
tool asks on the terminal (non-interactive runs deny them). --skill NAME runs a skill
explicitly, --cwd DIR selects the project root for skill discovery, and {config_dir}/.env
is loaded automatically. In the REPL, each turn streams its answer and keeps the conversation
for follow-ups.
FindSkills / ActivateSkill /
ReadSkillFilechat_structured() with retries and usage accountingLlmCompactor, and a full-history escape hatch, all inside the run deadlineAgentCore.chat(), .stream(), and .run_skill() with no CLI
dependenciespackages/cliver-core/ # embeddable library (published)
packages/cliver/ # CLI application built on cliver-core
website/ # documentation site (Astro + Starlight)
Full documentation at cliver-project.github.io/CLIver
git clone https://github.com/cliver-project/CLIver.git
cd CLIver
make init # Set up dev environment
make test # Run the workspace test suite
make lint # Lint and format check
make format # Auto-fix lint/format issues
make dist # Build the cliver-core sdist + wheel
cd website && npm install && npm run dev # Docs dev server at localhost:4321