Sub-agent hive-mind: a cross-tool shared-context MCP server. Every agent your tools spawn (Claude Code, Hermes, Cursor, Codex) shares one live context — semantic memory, a dependency-aware auto-scheduling task DAG, artifacts, roles, real-time push, a web dashboard, and crash recovery.
Subhive demonstrates solid tool design with 25 well-named, verb-prefixed tools (memory_*, task_*, artifact_*, etc.). All tools have descriptions (avg ~150 chars, within baseline 194). Input schemas are present and typed for all tools. However, output schemas are entirely undocumented, no tool specifies what fields it returns, forcing LLMs to infer structure. Error handling is minimal; most tools lack recovery guidance. Parameter descriptions are adequate but lack format constraints (e.g., task priority enum is declared but no min/max for numeric params like k in memory_search). Security is well-gated (admin-only tools marked, roles enforced), but descriptions don't explicitly state permission requirements. Composition is excellent, tools chain well (task_claim_next → task_update), and idempotency is implicit in most operations.
Retrieve a shared artifact by id or name (latest version for that name).
List all artifacts in this context (metadata only).
Store a shared artifact (code, plan, JSON, md, sql). Same name = new version (latest wins on name lookup).
[admin] Run the scheduler: distribute all unassigned ready tasks across the least-busy online agents (load-balanced). Returns the assignments made. Use to fan work out automatically instead of hand-assigning.
[admin] Create a new isolated context (workspace). Optionally set a token to require authentication. Returns the new context.
Get a compact human-readable briefing of the whole context — shared decisions, progress, in-flight work, what is ready to pick up, blockers, and a suggested next action. Call this right after joining to catch up on a project already in progress without paging the raw event log.
Output schemas completely undocumented. No tool specifies what fields it returns (e.g., roster returns agent list with status/role/load, but this is only in description text, not a structured schema). LLMs cannot plan downstream tool calls or extract fields reliably.
Error handling lacks recovery guidance. Tools do not specify what errors are retryable, user-fixable, or fatal. No tool returns actionable error messages (e.g., 'Task not found. Try task_list() to see available tasks').
Numeric parameter constraints missing. memory_search accepts 'k' (max results) with no min/max bounds documented. task_create priority enum is declared but no guidance on when to use each level. get_events 'since_id' lacks range documentation.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 68 | <=2025-11-25 | v2 |
Fetch the event log for this context, optionally filtered by type and/or actor. Useful for polling if you don't have WebSocket. Returns events in creation order (oldest first).
Read your direct messages (and optionally broadcasts). Returns messages sent to you since you last called inbox, newest first.
Delete a shared-memory entry by key.
Read a single shared-memory value by exact key. Returns null if the key does not exist.
List all shared-memory entries in this context, optionally filtered by a tag.
Semantic search over shared memory: find relevant facts by meaning, not exact key. e.g. query "how do we handle auth" returns the JWT decision even if the key is "auth-strategy". Returns top-k results ranked by similarity score. Fully local, no external API.
Store a key/value fact in shared context memory so every agent can read it. Overwrites an existing key. Use tags to group facts. Automatically indexed for semantic search.
[admin] Manually reap stale agents (those whose heartbeat is older than staleMs). Normally runs automatically every 60s.
List all agents connected to this context with online/offline status, role, and current task load.
Message another agent (set `to`) or broadcast (broadcast=true). Delivered in real time + readable via inbox.
[admin] Change an agent's role (reader/writer/admin). Only admins can grant admin.
Get a complete snapshot of the context: agents, memory, tasks, artifacts, messages, and the event log. Useful for dashboards or late-join catch-up.
Assign a specific task to yourself and mark it in_progress.
Atomically claim the single highest-priority ready task for yourself (assigns it to you + marks in_progress) and return it. Race-safe: two agents calling this concurrently never grab the same task. Returns null if nothing is ready. This is the one-shot "give me work" call.
Create a work item. Optionally assign it and declare dependencies via `needs` (task ids that must be done first). A task with unmet deps is "blocked" and hidden from task_next until they complete. Priority high/normal/low affects ordering.
List tasks in this context. Filter by status and/or assignee. Each task shows its dependencies and blocked state.
Return ready-to-work tasks: pending tasks whose dependencies are all satisfied, ordered by priority. Pass mine=true to only get tasks assigned to you or unassigned.
Update a task: status (pending/in_progress/done/cancelled), reassign, set result, or change priority. Marking "done" auto-unblocks dependents and emits task.unblocked.
Return your own identity in the hive: agent name, role, context, session id, and the other agents connected to the same context.
Permission requirements not declared in tool descriptions. Tools like auto_assign, set_role, context_create, reap are marked [admin] in description text only, no formal scope/permission metadata. Agents cannot determine least-privilege configuration.
Pagination not implemented for list tools. task_list, memory_list, artifact_list, get_events return all results with no limit/offset/cursor. Large contexts could return thousands of items, exhausting context window.