Claude Code

Host entry

A terminal command owns the process, current working directory, stdin/stdout, shell environment, and project file discovery.

Host owns

The CLI host owns agent loop timing, approval prompts, tool execution, session files, and how hook or MCP subprocesses are launched.

Connector role

agent-connector can only write the host's documented CLI config, prompt, memory, hook, MCP, or package files and point them at the home binary. Current wired connector surfaces: MCP server, Hooks, Commands, Skills, Subagents, Memory, Statusline handler.

Implementation consequence

A CLI surface is byte-addressable and smoke-testable from a temporary HOME, but terminal UI behavior still belongs to the host process. No visible host-native gap is currently tracked for this surface.

Source pass

Current source review findings are kept as the first architecture section so readers can separate external evidence from the generated local coverage model.

• The hook system is a native lifecycle layer, not a generic MCP callback: command, HTTP, prompt, agent, and MCP-tool handlers can be attached to named lifecycle events.
• Hook location is scoped: user settings, project settings, local project settings, managed policy, plugin hooks, and component frontmatter are distinct placement surfaces.
• The event vocabulary includes session, prompt, tool, subagent, task, compaction, worktree, file, configuration, and MCP elicitation events, which explains why this host is the reference rich json-stdio adapter.

Runtime boundary

The adapter target is the local Claude Code CLI session. Claude Code owns the model loop, tool execution, permissions, and session state; agent-connector installs into the host's native configuration, command, plugin, memory, and agent file surfaces rather than wrapping the whole CLI.

• The public repository describes Claude Code as a terminal agentic coding tool, while also pointing users to IDE and GitHub usage.
• Repository-visible `.claude/commands` and `.claude-plugin` paths make content and plugin surfaces concrete instead of only conceptual.
• The archive keeps Claude Code as a CLI row because the byte-addressable adapter target is the local command-line host.

MCP registration

Claude Code's MCP surface has local, project, user, plugin-provided, and connector scopes. agent-connector must preserve that scope model instead of treating MCP as a single global JSON file.

• Project-scoped MCP servers are shareable through project-root `.mcp.json`.
• User-scoped MCP servers live in `~/.claude.json` and are private to the user account.
• Scope precedence matters because duplicate server names are resolved by source priority, not merged field-by-field.

Hook bridge

Hooks are Claude Code's primary lifecycle interception layer. Command hooks receive event JSON on stdin, while HTTP, prompt, agent, and MCP-tool handlers expand the same lifecycle model beyond simple shell commands.

• Events span session, prompt, tool, subagent, task, compaction, worktree, file, configuration, and elicitation surfaces.
• Hook placement is multi-scope: user settings, project settings, local project settings, managed policy, plugin hooks, and component frontmatter can all matter.
• Decision control is event-specific, so adapter claims need to stay tied to event envelopes rather than a generic allow/deny model.

Content surfaces

Claude Code content is not one directory. Slash commands, skills, subagents, plugins, and memory each have separate file contracts and precedence rules.

• Subagents are Markdown files with YAML frontmatter in project, user, managed, CLI, or plugin scopes.
• Project and user subagents are recursively scanned, while plugin subagents use plugin-scoped identifiers.
• Plugin subagents intentionally ignore some powerful frontmatter fields such as hooks, MCP servers, and permission mode.

Marketplace and host-only affordances

Claude Code is a plugin-marketplace shaped host, not just a hook target. Plugin-provided MCP servers, commands, agents, and hooks are distribution surfaces, while statusline remains a host-native CLI affordance.

• agent-connector should model plugin delivery separately from runtime power: installing a plugin can add MCP, commands, agents, or hooks.
• Statusline support is host-owned UI state, so it remains separate from hooks and command content in the page inventory.

Host-specific shape

Claude Code is the reference rich json-stdio host: settings hooks, MCP servers, slash commands, skills, subagents, memory, plugin-provided MCP, and local plugin marketplace all exist as native concepts.

• Form factor: CLI.
• Native surfaces: MCP server, Hooks, Commands, Skills, Subagents, Memory, Statusline handler.
• Wired surfaces: MCP server, Hooks, Commands, Skills, Subagents, Memory, Statusline handler.
• Visible gaps: none.

Hooks

Hooks enter through a host-launched command that exchanges event JSON over stdio.

MCP

The adapter writes this host's native MCP registration dialect, including its root key, scope, transport fields, and environment syntax.

Memory

Managed memory text is written into the rules or memory file this host actually reads, with ownership markers for reversible uninstall.

Marketplace and affordances

A marketplace or plugin-driver path is wired, so install/update/uninstall can be driven rather than only emitted as files.

• Runtime handlers: Statusline handler.
• User-visible affordances: Users see managed commands + skills + subagents + statusline + memory surfaces after install; runtime-only surfaces stay behind the host boundary..