MCP server de clai
Commit 14 nov 2026Organiza y controla equipos de agentes de IA en entornos de sandbox utilizando herramientas nativas de MCP y ejecuciones programadas.
También conocido como clairun/clai
- Stars
- 9
- Herramientas
- —
- Descargas, 30 días
- —
- Última versión
- —
- Último commit
- 14 nov 2026
- Licencia
- MIT
¿Qué es clai MCP?
Organiza y controla equipos de agentes de IA en entornos de sandbox utilizando herramientas nativas de MCP y ejecuciones programadas.
Resumen generado a partir del README del proyecto
Acerca del MCP server de clai
Gestionar equipos especializados de agentes de IA dentro de entornos seguros y aislados (sandboxed) es lo que clai permite a los desarrolladores y a los flujos de trabajo automatizados. Los usuarios pueden organizar el trabajo en workspaces donde un agente principal supervisa a agentes de apoyo, como un Code Reviewer o un SoW Tracker, cada uno con sus propias herramientas, proveedores y permisos. Esta plataforma local-first permite la ejecución en paralelo y ejecuciones programadas, ofreciendo una vista de Fleet para supervisar todas las tareas activas y los problemas detectados.
La configuración implica conectar proveedores de API como Anthropic o endpoints compatibles con OpenAI, o dirigir agentes de CLI locales como Claude Code y OpenCode. Los desarrolladores pueden conectar MCP servers mediante transportes HTTP o stdio para ampliar las capacidades de los agentes. La seguridad se gestiona mediante permisos granulares del sistema de archivos y tres modos de shell: Off, Ask first y Don't ask. El acceso a la red se controla de forma similar mediante los modos Off, Allowlist o Host network, garantizando que los agentes operen dentro de los límites definidos mientras realizan tareas complejas.
Qué puedes hacer con él
- 1Tareas de agentes periódicas programadasPuedes configurar workspaces para que se ejecuten automáticamente en un intervalo establecido para realizar tareas de mantenimiento o monitoreo recurrentes.
- 2Ejecución segura de shell localEjecuta comandos de agentes con controles de seguridad estrictos que requieren aprobación manual para operaciones sensibles como git pushes.
Escrito a partir del README del proyecto. Cada caso de uso se verifica con una cita del mismo.
Verificaciones automáticas
Ejecutado 14 nov 2026Informe completo →Aún no verificado. Un servidor necesita una puntuación de 80, ningún hallazgo grave y suficientes datos para que podamos revisarlo. Cómo funcionan las verificaciones · Qué corregir para obtener más puntos
- Repositorio de origen vinculado y accesibleclairun/clai is public and reachablegithub.com/clairun/clai ↗Aprobado
- El paquete indica este repositorio como su origenNo verificado
- Procedencia de la compilación (atestación de Sigstore)No verificado
- Licencia de código abierto (OSI)MITspdx.org/licenses/MIT.html ↗Aprobado
- Commit o lanzamiento en el último añoLast commit or release 2026-11-14github.com/clairun/clai ↗Aprobado
- OpenSSF ScorecardNo verificado
- El nombre del paquete no es similar a un paquete popularNo verificado
- No hay avisos críticos en el paquete ni en sus dependenciasNo verificado
- Sin instrucciones ocultas para el modelo0 tool descriptions and the README read; no override phrases, credential paths or hidden charactersAprobado
- Las herramientas destructivas están etiquetadasNo verificado
- El endpoint remoto responde al handshake de MCPNo verificado
README
Ver en GitHubWhat it is
CLAI organizes work into workspaces. A workspace is an ongoing conversation with a main agent — a configurable LLM that owns the workspace's provider, tools, skills, and permissions. The main agent can:
- Chat with you in a surface that stays front and center
- Use tools through attached MCP servers and a local shell sandbox
- Delegate to helper agents in the same workspace, each with their own skills, tools, and execution policy
- Persist context — memories and artifacts it writes are inspectable from the workspace header
- Run on a schedule — periodic workspaces fire on their interval
The Fleet view supervises everything: scheduled workspaces float to the top, anything needing attention (failed or blocked tasks, input prompts) is flagged, and selecting a card slides in a live chat preview.
Features
- Workspace-local agent teams — Add helpers (e.g. a Code Reviewer or SoW Tracker) with their own prompts, skills, MCP servers, providers, and execution policy. The main agent calls them as tools, and you can read their full transcripts.
- Multiple providers — API connections (OpenAI-compatible or Anthropic-compatible) and local CLI agents such as Claude Code, OpenAI Codex, and OpenCode. Each agent picks its own.
- MCP-native tools — Configure MCP servers once in Settings, then attach them per workspace or per agent. HTTP and stdio transports.
- Local execution sandbox — Per-agent filesystem grants and three shell
modes: Off (no generic local file access), Ask first (every command
runs, except that commands on the agent's ask-first list, such as
git push,rmornpm publish, wait for your approval), and Don't ask (every command runs without asking). The ask-first list catches common accidents, not a determined agent: on Linux and macOS the OS sandbox and your path grants are the boundary. On an ask-first card, Always allow removes the rule from the agent's list: for a shared teammate in every workspace that uses it, for a workspace's main agent in that workspace only. Add a rule back in the agent's settings to be asked again. - Network modes — Each agent has one of three, in its Local access settings: Off (no network for shell commands or the web tools), Allowlist (only the hosts matching the agent's allowed hosts, about 100 package registries and code hosts by default, never those matching its denied hosts; web search still reaches its search provider), and Host network (unfiltered), which reaches any host, including services on your computer and your local network. Allowlist never connects to loopback or link-local addresses, so services that listen only on your computer's localhost stay out of reach, but a host you list that resolves to a local-network address is reachable from shell commands. In Allowlist an agent that needs another host asks for it, and you allow it for the run or always, allow its parent domain instead, or deny it.
- A shared agent library — Teammates are defined once in Settings → Agents and added to the workspaces that need them. Editing one reaches every workspace using it; each workspace still owns its own main agent, history and schedule.
- Default skills — CLAI registers the read-only
clairun/clai-skillsrepository by default. - Inspectable tasks — Delegated work streams a live transcript: the helper agent's full conversation, tool calls, and verdict.
- Memory & artifacts — Agents persist findings to the workspace directory; the drawer surfaces both, with read-only previews (rendered markdown, pretty JSON, multi-file HTML).
- Run notices — A run that hits a policy denial finishes in an amber "warnings" state and surfaces in the Fleet, instead of failing silently.
Local execution sandboxing is platform-specific. On Linux, shell commands run through bubblewrap; in Allowlist each command gets its own network namespace whose only way out is CLAI's network proxy, through the small
clai-sandbox-helperbinary installed next toclai. Tools that ignoreHTTP(S)_PROXY, and DNS lookups, fail there; choose Host network for them. On macOS, they run through Seatbelt viasandbox-exec, with Off or Host network only: Allowlist is not available on macOS yet. On Windows there is no sandbox backend: shell execution is labeled as a host shell, and while CLAI still validates the command's working directory against the filesystem grants, nothing stops the command itself from reaching outside them. Windows also ships no POSIX shell, sobash_execlooks for Git Bash, then MSYS2, thenbashonPATH, and fails with an "install Git for Windows" notice when it finds none. The allow/block lists are enforced by CLAI before a command spawns, on every platform. On Windows the network setting applies to the web tools only: Off and Host network restrict or open web fetch and web search, and shell commands use the host network. Allowlist cannot be chosen there; a stored one (a hand edit or a config copied from Linux) stops shell commands until you choose Off or Host network.
Install
Download the latest build for your platform from the Releases page:
| Platform | Download |
|---|---|
| Windows | .msi / .exe |
| macOS | .dmg |
| Linux | .deb, .rpm, or .flatpak |
| Arch Linux | AUR: clai-desktop-bin |
On Debian/Ubuntu you can install from the CLAI apt repository instead —
updates then arrive through apt upgrade like any other package:
sudo curl -fsSLo /usr/share/keyrings/clai-archive-keyring.gpg \
https://download.clai.run/apt/clai-archive-keyring.gpg
echo "deb [arch=amd64 signed-by=/usr/share/keyrings/clai-archive-keyring.gpg] https://download.clai.run/apt stable main" | \
sudo tee /etc/apt/sources.list.d/clai.list
sudo apt update && sudo apt install clai
On Fedora/RHEL, the CLAI rpm repository does the same for dnf upgrade:
sudo curl -fsSLo /etc/yum.repos.d/clai.repo https://download.clai.run/rpm/clai.repo
sudo dnf install clai
openSUSE: sudo zypper ar -f https://download.clai.run/rpm clai && sudo zypper install clai
(Installing the .deb/.rpm from the Releases page enrolls the same
repository automatically; see packaging/linux-repo/README.md, including
the opt-outs.)
On macOS, the builds are not yet signed or notarized with an Apple
Developer ID, so after downloading macOS may say "clai is damaged and
can't be opened". The app is not damaged. Move clai.app to
Applications, run this once in Terminal to remove the download
quarantine flag, then open it normally:
xattr -cr /Applications/clai.app
Getting started
- Add a provider — In Settings, connect an API provider or point CLAI at a local CLI agent. Supported CLI agents are auto-detected and pre-wired into new workspaces.
- Add MCP servers (optional) — Register local or remote MCP servers for external tools.
- Create a workspace — From the Fleet view. Open its settings (gear icon in the header) to attach a provider, MCP servers, and skills.
- Chat with the main agent.
- Add helper agents (optional) — Define them once in Settings → Agents, then add them to a workspace's team from its settings; the main agent can then delegate to them.
- Make it periodic (optional) — Toggle Schedule to run the main agent on an interval.
Providers
- API connections — Add an OpenAI-compatible or Anthropic-compatible provider with an API key and optional custom base URL. Works with OpenAI, Anthropic, together.ai, Groq, and local endpoints (vLLM, llama.cpp, Ollama).
- CLI agents — Drive a locally installed coding CLI directly, such as Claude Code, OpenAI Codex, or OpenCode. CLAI exposes its tools to them over MCP and streams their output like any other run.
Each agent chooses its own provider, so you can mix — for example a Claude Code main agent with an OpenAI-compatible reviewer.
Development
git clone https://github.com/clairun/clai.git
cd clai
npm install
make dev # run the desktop app in development
make dev builds the sandbox helper before it starts the app. cargo run and
npm run tauri:dev build only the app, and the Allowlist network mode also
needs the helper next to it: build it with
cargo build --manifest-path src-tauri/Cargo.toml --bins, or shell commands in
Allowlist fail with "network helper not found".
Before pushing:
npm run lint && npm run format:check && npm run build
cargo fmt --manifest-path src-tauri/Cargo.toml --check
cargo clippy --manifest-path src-tauri/Cargo.toml --all-targets -- -D warnings
cargo test --manifest-path src-tauri/Cargo.toml
Architecture
- Frontend — React + Tauri; a chat-first workspace with a drawer for agents / tasks / memories / artifacts and slide-out transcript and file panels.
- Runtime — Each workspace owns its main agent, its assignments of shared agents, and a persistent session. Built-in tools (shell execution, inter-agent calls, task management) plus MCP tools, gated by each agent's policy. An agent is resolved fresh at the start of every turn, so a shared edit lands on the next turn and never mid-tool-sequence.
- Scheduler — Periodic workspaces run from the agent runner, emitting the same streaming events as interactive chat.
- Skills — Discovered from read-only local or git sources. The app-managed
default source is
clairun/clai-skills.
License
MIT
Preguntas sobre clai MCP
¿Cómo gestiona CLAI la seguridad y el sandboxing?
CLAI utiliza sandboxing específico de la plataforma: bubblewrap en Linux, Seatbelt en macOS y permisos de sistema de archivos en Windows para controlar el acceso a la shell y a la red.
¿Puedo usar mis propios proveedores de LLM?
Sí, puedes conectar proveedores compatibles con OpenAI o Anthropic, incluyendo endpoints locales como Ollama o vLLM.
¿Pueden los agentes ejecutarse automáticamente con un temporizador?
Sí, puedes configurar workspaces para que se ejecuten según un horario mediante ejecuciones periódicas.
¿Es compatible con MCP servers?
Sí, puedes configurar MCP servers a través de transportes HTTP o stdio y asignarlos a agentes o workspaces específicos.