Agents

Overview

An agent is the AI coding tool that executes inside the sandbox. You send a prompt and a repository — the agent does the work. coSPEC handles the infrastructure, isolation, and output delivery.

Currently, coSPEC supports Claude Code — Anthropic's CLI coding agent. See the Claude Code page for setup and agent-specific configuration details.

Capabilities

Inside the sandbox, the agent can:

• Read, write, and edit files
• Run shell commands (bash)
• Use git (commit, branch, push)
• Create GitHub PRs and issues
• Search the web
• Spawn sub-agents for parallel work

All actions are constrained by the sandbox's egress rules and the run's guardrails.

MCP Support

MCP (Model Context Protocol) is an open standard that lets AI agents connect to external tools and services — databases, APIs, project management tools, and more.

Agents running in coSPEC can use MCP servers that are configured in your repository. When a run starts, the agent discovers the MCP configuration and connects to the servers automatically. See the Claude Code page for how to configure MCP servers in your repo.

Skills

Skills are reusable commands and workflows that extend what an agent can do. They let you define project-specific tasks — linting rules, deployment scripts, custom code generation patterns, or any repeatable workflow.

Skills are installed in your repository. When a run starts, the agent picks them up automatically and can invoke them during execution. See the Claude Code page for how to configure skills in your repo.

Project Instructions

Project instructions give the agent context about your codebase before it starts working — coding conventions, architecture decisions, build commands, and anything else relevant to your project.

Instructions are stored as files in your repository. When a run starts, the agent reads them automatically. See the Claude Code page for how to set up project instructions in your repo.

Supported Agents

More agents are on the roadmap. See Closed Beta for details.