SlopCamera
Browse documentationTutorials · Set up SlopCamera for Cursor, Devin CLI, and other coding agents

Tutorials

Set up SlopCamera for Cursor, Devin CLI, and other coding agents

Install the Agent Skill where Cursor and Devin CLI load it, run it as /slopcamera, and add the MCP server or plain CLI for any other agent.

Cursor, Devin CLI, and other agents without a dedicated SlopCamera guide have two routes in: the portable Agent Skill directory for agents that read the agents convention, and the plain slopcamera command for agents that only run shell commands. Both come from the same install. If the client speaks MCP, its fixed tools cover diagrams, images, and scene inspection and planning; see Use SlopCamera from an MCP client.

Install the portable skill

slopcamera skill install --target agents

The agents target is the generic one. A user-scope install writes ~/.agents/skills/slopcamera; adding --scope project inside a repository writes .agents/skills/slopcamera there instead, and --project <directory> names a different root. The folder is self-contained: SKILL.md, its task references, and agent metadata travel together, so an agent that reads the portable Agent Skills layout gets the same version-matched guidance as the Codex and Claude Code targets.

To place the guidance somewhere else, such as a custom instructions directory or a committed agent file, slopcamera skill path prints the bundled directory to copy from. Installing a skill never installs the CLI or the optional native engines.

Use the skill in Cursor or Devin CLI

Cursor and Devin CLI both load skills from ~/.agents/skills and from a project's .agents/skills, so the agents install above works in both. Restart Cursor, or start a new Devin CLI session, after installing.

SetupCursorDevin CLI
Folders it loads a SlopCamera install from.agents/skills and ~/.agents/skills, plus .claude/skills, .codex/skills, and their home-directory forms.agents/skills and ~/.agents/skills, plus .claude/skills and ~/.claude/skills
Run the skill yourselfType / in Agent chat and choose slopcameraType /slopcamera
Add the MCP serverPut the mcpServers entry from the MCP tutorial in ~/.cursor/mcp.json, or in .cursor/mcp.json for one projectRun devin mcp add slopcamera -- slopcamera mcp --root /absolute/path/to/workspace

Both agents can also load the skill on their own when a request matches its description. devin mcp add saves the server for the current project in .devin/mcp_config.local.json; add -s user to save it in ~/.config/devin/mcp_config.json for every project, and run devin mcp list to see the servers Devin CLI has configured. Devin CLI also lists servers from a project's .mcp.json and .cursor/mcp.json, so a project that already registers SlopCamera for Claude Code or Cursor needs no second entry.

Keep one of these skill folders. With both ~/.agents/skills/slopcamera and ~/.claude/skills/slopcamera present, Devin CLI lists two copies, /agents:slopcamera and /claude:slopcamera. If you also use Claude Code, install only with slopcamera skill install --target claude; Devin CLI and Cursor read that folder too.

The Devin CLI folders and MCP commands were checked on 4 October 2026 with Devin CLI 3000.11.3 and SlopCamera 3.10.3: devin skills list showed /slopcamera from each folder in the table but not from ~/.codex/skills, and devin mcp list showed the added server. No Devin CLI session ran the skill. Devin documents its own folders in its skills overview and the Claude Code and Cursor files it reads in configuration import. The Cursor column comes from Cursor's skills and MCP documentation, read the same day; this setup was not tested in Cursor.

Drive the CLI directly

An agent that only runs shell commands needs the slopcamera command on its PATH and a few lines of direction in its instruction file. Point it at the discovery commands rather than a frozen command list:

slopcamera --help
slopcamera help <family>
slopcamera doctor --json
slopcamera operations list --json
slopcamera workflows list --json

Every read and mutation accepts --json for machine-readable output, and slopcamera code search plus slopcamera code execute expose the fixed portable operation registry for programmatic calls. If the agent reads AGENTS.md in a repository, a short paragraph there is enough: name the slopcamera command, require the discovery commands above before assuming an option exists, and ask the agent to keep authored sources editable beside derived media and to report real output paths.

Use the SDK

An agent embedding TypeScript can skip the shell and import the same contracts: @hraness/slopcamera for the portable SDK, @hraness/slopcamera/code for declarative graphs, and @hraness/slopcamera/workflow for trusted Bun workflow modules. The local subpaths ship with the source-backed Bun package or a checkout. SDK surfaces describes what each entrypoint includes.

What the agent still needs

  • Generation needs the caller's AI_GATEWAY_API_KEY or VERCEL_OIDC_TOKEN in the process environment; this site never receives either.
  • Native engines such as Blender, CadQuery, and Manim install separately, and running authored native source additionally needs the invocation-scoped --allow-trusted-code flag, which grants current-user execution without an operating-system sandbox.
  • The released CLI includes the scene, studio, and durable-run commands. The capability reference lists what each install includes and its platform requirements.