llm-tools

Experimental LLM agent environments configured via Nix flakes.

Agents

Claude Code

nix develop github:cbarber/llm-tools#claude-code

OpenCode

nix develop github:cbarber/llm-tools#opencode

Both agents include:

• Sandboxed execution (bubblewrap on Linux, sandbox-exec on macOS)
• cclsp MCP server for LSP integration
• Auto-detected language servers (Nix via nil)
• Beads task management with git integration
• Shared tooling (smart-lint, smart-test, notify)

Sandboxing

Agents run in restricted environments:

• Read/write: Project directory, temp workspace
• Read-only: /nix store
• Blocked: Home directory, other projects

See tools/SANDBOX.md for details.

Task Management

Beads provides git-backed issue tracking:

bd ready                    # Show available work
bd create "task" -p 2       # Create issue (P0-P4)
bd close <id>               # Complete work
bd sync                     # Sync with git

Auto-initializes in agent shells.

Configuration

Auto-generates on first run if missing:

• .mcp.json / opencode.json - MCP server configuration
• cclsp.json - Language server mappings for detected project languages

Set ANTHROPIC_API_KEY in .env for API auth. Set AUTO_LAUNCH=false to skip auto-launch.

Structure

• agents/ - Agent environments (claude-code, opencode)
• tools/ - Shared MCP servers (cclsp), scripts, and sandbox
• flake.nix - Root configuration