Related Servers

There is real overlap: claude_agent, codex_agent, and grok_agent are explicit legacy aliases of agent_launch, and claude_turn duplicates the result-recovery role of pty_read(agent_transcript:true), while agent_approval vs claude_approval and pty_list vs pty_observe are close cousins. The descriptions do flag the aliases and per-harness scoping clearly, so a careful agent can disambiguate, but the surface still invites misselection.

A clear prefix-by-resource convention holds throughout (pty_open/pty_read/pty_send/pty_key/pty_list/pty_close/pty_observe, agent_launch/agent_auth/agent_models/agent_approval/agent_goal/agent_configure, claude_turn/claude_approval/claude_agent). Minor deviations exist: harness-named legacy tools (codex_agent, grok_agent) and the bare noun 'diagnostics' and verbless 'claude_turn' break the verb_noun rhythm.

At 19 tools it sits in the heavy-but-borderline band; the core set (PTY lifecycle, agent launch, auth, models, configure, approvals, observe) is defensible for an agent-orchestration server spanning five harnesses. However, three of the 19 are deprecated compatibility aliases that inflate the surface without adding capability, plus claude_turn overlaps existing read paths.

Coverage of the domain is solid: session open/close/list/observe/read/send/key, agent launch, auth, model catalog, live reconfiguration, per-harness approvals, goal management, and diagnostics. Gaps are minor — no explicit resize, no agent-launch listing/status tool, and no dedicated completion-wait tool despite wait_process being referenced repeatedly.