Hythe MCP server

Read from the package source without running it. The installed server may list more.

• add_observations

  Add new observations to existing entities with automatic vector embedding and graph updates. Supports mode:"replace-current" (alias supersedesLatest:true) to atomically supersede the entity's current observation server-side — the recommended way to maintain a one-current-observation entity without fetching the prior observation id.

• archive_messages

  Archive messages for an agent — either specific messageIds, or all messages older than N days. Set markAsRead:true to acknowledge and archive the same scoped messages atomically. Archived messages are excluded from get_ai_messages by default.

• begin_session

  Compatibility wrapper over resume (Engram v1 B2, TOOL-COMPATIBILITY-MAP.md). Rebuilds working context for a scope; NEVER auto-consumes handoffs — acknowledgement is explicit via ackHandoffIds. Accepts the wrapper shape {agentId, scope:{project|task}, budget, ackHandoffIds?} or the pinned legacy shape {agentId, projectId, maxTokens?, userId?} (auto-adapted: projectId→scope.project, maxTokens→budget

• checkpointwrite action

  Write an immutable, branch-preserving state snapshot for a scope (CAS on expectedRevision; stale parents branch, never conflict) with fingerprint-verified idempotency, plus transactional fact/loop changes. Returns the ids each factChange/loopChange materialized to (result.changes, positional, with a created flag) so callers never need a follow-up resume to learn a new loopId/factId. The counterpar

• compact_memory

  Analyze and reclaim storage bloat (ENG-2). Classes: "index-diet" (rebuild the derived graph_lookup_keys index under the current extraction policy), "superseded" (retire explicitly-superseded observations to restorable trash — marked-only, never guesses staleness, never touches an entity's current observation), "vec-orphans" (drop vector rows whose source memory is gone), "message-archive" (flag re

• create_entitieswrite action

  Create multiple new entities in the knowledge graph

• create_relationswrite action

  Create multiple new relations between entities in the knowledge graph

• delete_entitywrite action

  Delete an entity and cascade-delete its observations and relations from the knowledge graph. Requires admin-equivalent authorization.

• delete_observations_by_entitywrite action

  Delete ALL observations for a given entity. Requires admin-equivalent authorization.

• discover_related_context

  Read-only related-knowledge discovery for an exact project/task scope. Automatically retrieves vector candidates, reranks them with bounded one/two-hop graph paths, and explains currentness, evidence references, provenance, and score components. It never creates relations or changes memory; use create_relations separately only after validating a proposed relationship.

• end_session

  Compatibility wrapper over checkpoint (Engram v1 B2, TOOL-COMPATIBILITY-MAP.md) — end_session IS checkpoint: pass {agentId, scope:{project|task}, expectedRevision, idempotencyKey, state, factChanges?, loopChanges?}. The legacy {projectId, summary, openItems} shape is RETIRED and returns a migration error. Legacy side effects are gone — the wrapper only checkpoints. Prefer checkpoint for new caller

• gc_agent_registrations

  Dry-run-first garbage collection for expired or stale inactive agent registrations. Deletes only when dryRun is false.

• get_agent_context

  Retrieve a tiered context bundle for an agent, optionally scoped to a project. Returns identity, project state, unread messages, guardrails, and handoff flags.

• get_agent_status

  Get comprehensive status and health information for AI agents

• get_ai_messages

  Retrieve messages for the calling agent's exact inbox with filtering and real offset pagination. Identity-bound client bridges require agentId to match the configured lane identity. The response reports the true total matching the filters plus hasMore/nextOffset. IMPORTANT: For routine inbox checks, just pass agentId — the defaults (unreadOnly: true, compact: true, limit: 5, offset: 0) return the

• get_current_observation

  Return the single authoritative (current) observation for an entity: the newest observation not superseded by any other. Resolves the supersedes chain server-side via the entity lookup index — no history scan, works for entities with thousands of observations. Returns id, timestamp, kind, canonicalFact, full contents, and metadata; current:null when the entity has no observations.

• get_entity_backlinks

  Bounded backlink view for ONE entity. Returns relation rows that point TO the entity, plus optional outgoing links for comparison. Inspired by Obsidian backlinks, but read from the SQLite knowledge graph.

• get_entity_detail

  Retrieve full entity content by storage ID, canonical entity name, or alias. `ids` preserves the scan-then-detail workflow; `names` or the singular `entity` input skips the search-then-refetch round trip and returns resolution metadata. STATE SEMANTICS: an entity's embedded observations[] is its creation-time definition snapshot, NOT current state — each entity result carries a resolved `currentOb

• get_entity_neighborhood

  Bounded local-graph view around ONE entity: the entity plus its directly-related entities and the relations among them, out to a small depth (1-2 hops). The safe, focused alternative to read_graph — hard-capped node/edge counts with truncated flags. Use it to answer "what is connected to X?".

• get_individual_memory

  Retrieve an agent's individual memory snapshot

• get_message_detail

  Retrieve the full content of a single message by ID. Use after scanning compact message list from get_ai_messages.

• get_user_profile

  Get a user profile within the current tenant. Returns timezone, locale, preferences.

• mark_messages_read

  Mark specific messages or all unread messages for an agent as read

• read_graph

  Read the knowledge graph in BOUNDED pages (entities, relations, and optionally observations). Defaults to a capped page and EXCLUDES observations unless requested, so a broad read can never dump the whole graph. Use limit/offset to paginate and since to fetch only recent rows; the response includes accurate totals + nextOffset. For a focused view around one entity, prefer get_entity_neighborhood.

• record_learning

  Record a learning entry into an agent's individual memory

• register_agent

  Register or refresh an AI agent in the collaboration system. Optional TTL/expiresAt metadata lets ephemeral agents expire without schema migration.

• remove_observationswrite action

  Remove specific observations from an entity by ID or content match. Requires admin-equivalent authorization.

• resume

  Rebuild working context for a project/task scope: current state, open loops, scoped messages/handoffs, facts, decisions, evidence handles and pointers — with closed per-section coverage accounting under a hard token budget. Pass resultVersion:2 to also receive `capsule` — the scope entity's rehydration capsule selected by kind (newest unsuperseded kind=capsule observation, with other unsuperseded

• search_entities

  Advanced federated search across graph, vectors, and cache. AGENT TIP: for a KNOWN entity/project name, pass searchType=exact — it is fast, precise, and name-anchored; hybrid/semantic is a bounded, lower-precision supplement best for fuzzy/exploratory queries (entity & observation hits are ranked above ai_message chatter). Returns compact summaries by default (use get_entity_detail for full conten

• search_nodes

  DEPRECATED alias for graph-only search; use search_entities with { searchType: "graph" }

• send_ai_messagewrite action

  Send messages to AI agents (direct, capability-based, or broadcast)

• set_agent_identity

  Temporarily disabled during migration to tenant-scoped stable principals. Calls fail without changing identity history; register an exact handle with register_agent instead.

• set_preferenceswrite action

  Update agent preferences in individual memory

• unregister_agent

  Soft-unregister an agent registration by marking it inactive and recording lifecycle audit metadata. Does not delete rows.

• update_observationwrite action

  Update the content of a specific observation. Runs sanitizer for parity with create/add paths. Requires admin-equivalent authorization.

• update_user_profilewrite action

  Update a user profile. Only the profile owner (or a service key) can update. Bumps prefs_version.

• Code scan124 source files scanned20/25
• –Live reliabilityno gateway calls yet and no remote to proben/a
• –Tool poisoningtools not inspected (local package is not executed); not countedn/a
• Auth qualitystatic API keys via environment variables6/15
• Maintenancelast push 15 days ago15/15
• Maintainer identitynamespace and repository owner differ4/10

Hythe: common questions

Is Hythe MCP server safe?

With care: it is graded C, so read the findings first (69/100). Read the Hythe safety report

How do I install Hythe?

It runs on your machine. Copy the Claude Code, Claude Desktop or Cursor config from the install section.

Does Hythe need an API key?

Yes. The registry entry asks for API_KEY.

Is Hythe maintained?

The last commit was 16 days ago (2026-09-05). The latest release is v0.2.0.

What can I use instead of Hythe?

Servers from other publishers that do the same job: Hicortex — AI Fleet Memory MCP server, AISIX AI Gateway MCP server and HOL Guard MCP server. Compare all Hythe alternatives.