Loading ToolBench
Exposes agent-manager's session commands as MCP tools over stdio, so any MCP-capable agent discovers and calls them natively. The manager registers this server into every session it spawns; the session id travels via the AGENT_MANAGER_SESSION_ID environment variable.
The server provides 25 tools with varying quality. All tools have names starting with action verbs and all include descriptions (10-100+ chars). Input schemas are fully visible and type-correct. However, descriptions are often terse (20-80 chars), some parameter descriptions lack actionable detail, and output schemas are not documented. Error handling guidance is absent. No tool annotations (readOnlyHint/destructiveHint/idempotentHint) despite clear risk metadata. Composition is solid, tools chain well and focus on discrete responsibilities (session management, terminal control, file reservations, code review). Parameter descriptions generally explain WHAT but not always FORMAT or CONSTRAINTS. The server follows 40-50% of the 54 agentic patterns; it lacks error recovery guidance, output schema docs, and constraint enforcement clarity.
Archive or restore a session
Close a terminal session
Create a new session group
Create a new agent session
Create a new terminal session
Delete a session group and move its sessions
Kill a running session
List all session groups
Output schemas are not documented. Tools return data but LLMs cannot plan downstream calls without knowing the response structure (field names, types, pagination info).
No tool annotations despite clear risk metadata. Tools marked DESTRUCTIVE (kill_session, delete_group) and WRITE-heavy lack readOnlyHint/destructiveHint annotations that would signal to LLM clients which operations are irreversible.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 65 | 2026-07-28+ | v2 |
List all active file reservations
List all agent sessions
List all terminals in the session group
Get the status of a message sent to a session
Read the current screen content of a session
Read the current screen content of a terminal
Release reserved file leases
Rename the session
Report a bug or request a feature
Reserve files or globs for exclusive editing
Open a code review for the session
Comment on a review point
Revive a stopped session
Send a message to another agent session
Send a command or keys to a terminal
Manage shared tasks (list, create, claim, finish, release, or delete)
Wait for a session to reach a specific state
Error handling provides no recovery guidance. Tool descriptions do not explain what errors are retryable, which require user intervention, or what the LLM should do next on failure.
Parameter descriptions lack format constraints. E.g., 'name' for rename expects 'kebab-case 2-4 words', this should be enforced via enum or pattern, not just prose. 'mode' in review accepts uncommitted|branch|last_commit|staged but is not declared as enum.
List tools lack pagination. list_terminals, list_sessions, list_groups, list_reservations have no limit/offset/cursor parameters; they return all results, risking context explosion for large deployments.
Descriptions are terse (40-80 chars on average). Many lack WHEN-to-use context, prerequisites, or next-step guidance. E.g., 'Revive a stopped session' does not explain what states qualify or what happens on success.
Composition issue: task tool is overloaded. It handles list, create, claim, finish, release, delete via an 'action' enum. This combines 6 concerns into one tool. Should split into task_list, task_create, task_claim, task_finish, task_release, task_delete for clearer intent.
No idempotency guidance. Tools like create_session, create_terminal, reserve_files modify state but do not indicate whether repeating the same call is safe or will cause duplicates/side effects.