Skip to main content
Glama
671,353 tools. Updated 2026-10-18 00:19

"Nexus Mods" matching MCP tools:

  • Search tracked product families by vendor, product line, or series name (case-insensitive substring, e.g. "nexus 9300"). Read-only. Returns up to 10 families with status, EOSL window, and page URL. Use this for discovery when you have a name; for an exact part number use lookup_part, and use get_family with a returned slug for the full record.
    ConnectorNo auth
  • Create a new entity in the world model. Always inserts a new node; it does not check for an existing entity with the same name. To reuse-or-create by name, use ``relate_entities`` (which also records a relation) or check with ``search_entities`` first. Side effects: with the default ``hybrid`` strategy the entity is linked to the NEXUS root entity (``belongs_to``); a creation episode, memory links and an event are recorded; and entities with a very similar name may be queued as merge proposals. For explicit links to specific entities use ``create_entity_cognitive``. Returns the entity plus ``relations_created``, ``episode_id``, ``event_id`` and ``memory_links_created``. Args: name: Human-readable name of the entity. entity_type: Type from the closed ontology (schema enum). Use 'unclassified' if none fits. attributes: Optional key-value attributes. strategy: Auto-linking strategy: 'hybrid' (default) or 'root' — link to the NEXUS root entity. 'similar' — link (related_to) to the most recently created entity of the same type. 'none' — no automatic links.
    ConnectorAPI key
  • Correct an existing entity's type, canonical name or aliases. Each change increments the entity's version. Changing the name or aliases re-indexes the entity and may queue merge proposals with similarly named entities. To combine two entities that are the same, use decide_merge_proposal rather than renaming one. At least one field must be given. Args: entity_id: UUID of the entity. entity_type: New type from the closed ontology. Rejected if any of the entity's current relations would become invalid for it. name: New canonical name. The old name is not kept automatically; include it in aliases to keep it searchable. aliases: The COMPLETE alias list (replaces the current one).
    Connector
    Destructive
    API key
  • Find entities by name or description, or filter them by type and attributes. With ``query``: returns ranked candidates, each with ``method`` (how it was found), ``score`` (0-1), ``legacy``, ``relaciones`` (counts by state) and ``contexto`` (up to 4 relations, current first, to tell similarly named entities apart). Methods, most to least reliable: exacta, alias, parcial (whole word or same words), semantica (name similarity) and memoria (entity mentioned in the closest episodes). semantica and memoria results appear only for multi-word queries and are candidates, not confirmed matches. A single word with no lexical match returns []. If nothing is found, the item may still exist in memory without being an entity (see retrieve_memories). Without ``query``: returns all entities matching entity_type and/or the given attribute values, with no result limit. With no query and no filters it returns every entity. Args: query: A name, part of a name, or a short description. entity_type: Filter by type (only without query). 'unclassified' means entities whose type is not known. attributes: Exact attribute key-value pairs to match (only without query). limit: Maximum results for a query search (1-50). via_memoria: With query, set False to skip episode-based matches and search names only.
    ConnectorAPI key
  • Create a directed relation between two existing entities in the world model. Both entities are referenced by UUID; to relate entities by name use `relate_entities`. Behavior: - `relation_type` must be a type from the closed ontology. When none fits, use 'related_to' and record the nuance with `store_fact`. - 'same_as' does not create a relation: it opens a merge proposal for the two entities, pending review, and returns its id. - Rejected with an error: a source or target that does not exist; a duplicate (a current relation of the same type between the same two entities); a relation from an entity to itself, or one that would close a cycle of current relations of the same type (the last two do not apply to 'related_to'). - For single-valued relation types, if the source already has a current relation of that type to a different target, the new relation is stored and a contradiction is opened (see `resolve_contradiction`). Args: source_entity_id: UUID of the source entity. target_entity_id: UUID of the target entity. relation_type: Relation type from the closed ontology (enum in the schema). weight: Strength of the relation. Default 1.0. causal_probability: Probability (0-1) that the source causes the target. Default 0.0. properties: Accepts only the keys 'weight' and 'causal_probability'. Any other key returns an error and no relation is created; store descriptive metadata as a fact with `store_fact`.
    ConnectorAPI key
  • Edit an existing goal's definition or progress. Only the fields you pass change; fields you pass replace the stored value entirely (lists are not merged). At least one field is required. To change the plan's steps, resubmit the plan with submit_goal_plan instead. Args: goal_id: UUID of the goal. progress: New progress value in [0.0, 1.0]. description: New description. priority: New priority on a 1-10 scale; stored as priority_weight 0.0-1.0 (priority=8 is returned as priority_weight 0.8). name: New name. success_conditions: Replaces the goal's done-criteria (target_state.conditions), which is what new plans are generated against. Example: {"test_coverage": ">=0.9", "description": "concrete, verifiable criterion"}. success_metric: How success is measured and evaluated. Set a numeric target and operator to use report_goal_measurement. Example: {"metric": "test_coverage", "target": 0.9, "operator": ">="}. constraints: List of constraints. Stored for reference only; NEXUS does not enforce them. linked_entities: List of entity UUIDs linked to the goal. time_horizon: Free-text time horizon, e.g. 'Q4 2026' or '2 weeks'.
    Connector
    Destructive
    API key

Matching MCP Servers

  • F
    license
    A
    quality
    A
    maintenance
    Exposes the Nexus Mods API to AI agents so they can search and inspect mods, browse authors' catalogues, and upload or publish mod files from a local archive in a few aggregated tool calls. It handles authentication, caching, reporting quotas, and the multi-step v3 upload flow, while gating write and publish operations behind opt-in permissions.
    22
    -
  • A
    license
    B
    quality
    C
    maintenance
    Enables MCP clients to browse the Nexus Mods catalogue and manage a local Vortex installation by reusing the user's existing Vortex login, including queuing download, install, enable/disable, deploy, and removal operations with job polling.
    21
    MIT

Matching MCP Connectors

  • Record a measured value for a goal's success metric and re-evaluate the goal against its target. Stores the value as success_metric.current, recomputes the goal's progress from it, and returns the evaluation: progress_score, confidence_score, success_met and calculation detail. It does not close the goal; use close_goal when success_met is true and you want to conclude it. The goal needs a success_metric with a 'target' and usually an 'operator' (default '=='), set with update_goal, e.g. {"metric": "test_coverage", "target": 0.9, "operator": ">="}. Without a target the call fails with 422. Goals evaluated by a structured success_condition also reject measurements with 422. Args: goal_id: UUID of the goal. value: The measured value. Zero is a valid measurement. baseline: Starting point used to compute progress, mainly for reduction goals. Overwrites any stored baseline. If omitted and none is stored, the previous measurement (or this one, if there is none) becomes the baseline.
    ConnectorAPI key
  • Record that a rule was checked against the current state of things (code, data, documentation) and still holds. Rules lose confidence over time and low-confidence rules are queued for review. Affirming resets the rule's staleness clock and restores its confidence to the configured maximum, taking it out of the review queue. It is meant for rules actually verified as correct, not for rules that were merely retrieved or look plausible. Args: rule_id: UUID of the verified rule.
    ConnectorAPI key
  • Send a text message to another AI agent in the user's workspace (for example Claude Code or Cursor). Messages are delivered to the recipient's inbox. Recipient (`to`): - an agent's label, a unique part of the label, or its principal_id (see list_agents); - '*' (or 'all'): broadcast to every agent that accepts messages from the sender. `skipped_not_accepting` reports the agents that were skipped; - 'me': another of the caller's own sessions. It requires `session_id`. To reply, pass reply_to=<message_id> and omit `to`. The reply goes to the sender in the same thread. Agents of the same person can always message each other. Other people's agents are reached only if their owner opted in (see messaging_preferences). Delivered does not mean read: the message waits in the inbox until it is read, or until it expires unread after 14 days by default. Args: message: Message text (max 20,000 characters). to: Recipient, as above. It may be empty only when reply_to is set. session_id: Optional target session. Required when to='me'. reply_to: Optional message_id being answered. my_session_id: Optional sender session ID, so replies can reach that session. ttl_hours: Unread lifetime in hours, 1-720. 0 means the default of 14 days.
    ConnectorAPI key
  • Return the optional setup kit for agents that run on the user's own machine (Claude Code, Cursor). The kit contains three hooks (recall at session start, recall on each prompt, sync at session end), the settings.json block, a usage policy, and an idempotent bash script that installs them. With the hooks installed, memory is recalled and stored automatically. Without them, NEXUS only has what the agent stores through its tools. Installing writes files and settings on the user's machine, so use this when the user asks to set NEXUS up, and let them run or approve the script. It does not apply to chat clients (Claude web or desktop, ChatGPT). Side effect: on an OAuth connection, the call issues the hooks credential and embeds it in the script. This works once, because a secret that was already shown is not returned again. On a plain `nxa_` token connection it does not issue a credential, and an account owner must issue one. `steps` says which case applies. Args: formato: "script" (default) for the bash installer plus steps, or "files" for the same files as text, for clients without bash. It is ignored for Cursor. client: "claude-code" (default) or "cursor". Cursor gets a session-start hook and a turn-recap hook. It has no per-prompt recall, because Cursor's beforeSubmitPrompt cannot inject context.
    ConnectorAPI key
  • Call this BEFORE an action that matters, and use the answer to decide. An action matters when at least one of these holds: someone else will receive or see it (an email, a message, a post, sharing a document); it commits the person (paying, quoting a price, accepting a meeting, signing); it cannot be undone (deleting, cancelling, filing, submitting); it touches something shared or official (a client record, the accounts, a government portal, the production system); it carries sensitive data; or it acts in the person's name. Reading, searching and drafting never need this call. Before calling, ask yourself: can this be undone? who does it reach? is there an agreed way to do this here? could there be a rule or a preference of the person about it? If any answer is "yes" or "I don't know", call with `consultar=True`. If you judge there is nothing to look up, still call with `consultar=False` to record your assessment: NEXUS will tell you anyway if it knows of a serious past mistake for this action. NEXUS returns the person's rules (with the mistake that created each one), their procedures, what to watch for, and a verdict: - "proceder": go ahead; - "proceder_con_lista": go ahead, applying the list; - "preguntar_a_la_persona": explain to the person what you will do and why, and wait for their confirmation; - "no_hacerlo": do not do it, and do not try another way to reach the same effect; tell the person what you wanted to do, only they can approve it. If you follow a returned procedure, report it with record_procedure_usage. Args: accion: What you are about to do, in your own words. intencion: Why: what you were asked and which part of the work this is. objetivo: Who or what it is done to (a client, a case, a patient, a channel, a project, a repository). riesgo_percibido: Your own answer to the four questions above. consultar: True to have NEXUS look up rules and procedures; False to only record that you judged it unnecessary. clave: The key given by a stop point (hook), if one stopped you. herramienta: The tool you are about to call, if any. detalle: Optional exact text (the email, the command). Stored truncated and with secrets masked, never whole. session_id: Your NEXUS session id, if you have one.
    ConnectorAPI key
  • Add an extra clickable shortcut (sitelink) that appears beneath a Google Ads ad — helpful deep-links like "See Pricing", "Book a Demo", or "Contact Sales" with two short description lines. Sitelinks lift click-through rate by giving searchers alternate landing paths to the same advertiser. ALSO KNOWN AS: sitelink, site link, extra link, ad link, secondary link, deep link, additional link, jump link KEYWORDS: sitelink, site link, link, deep link, shortcut, google ads, extension, pricing link, demo link, contact link, landing page, CTR, click through WHEN TO USE: - "Add a sitelink to our Google Ads library" - "Create a 'See Pricing' / 'Book a Demo' / 'Contact Sales' link extension" - "Register a new sitelink for the homepage redesign" - "I need a shortcut under our search ads that points to the new pricing page" WHEN NOT TO USE: - Want a non-clickable selling-point snippet ("Free Shipping", "24/7 Support") → use create_google_callout_extension - Want a labeled list of offerings (e.g. "Brands: Nest, Nexus") → use create_google_structured_snippet_extension - Same sitelink for Microsoft Ads / Bing → use create_microsoft_sitelink_extension (library entries are per-channel) - Anything for Facebook / LinkedIn / Reddit — ad extensions are a search-channel-only concept INPUTS (all required): - url: destination the sitelink opens. - link_text: visible link text (1-25 chars). - description1: first description line under the link (1-35 chars). - description2: second description line (1-35 chars). EXAMPLE: create_google_sitelink_extension( url="https://example.com/pricing", link_text="See Pricing", description1="Plans for every team size", description2="Start free, upgrade anytime", ) Returns the persisted extension (id, externalId, type=SITELINK, channel=GOOGLE_ADS). A 400 usually means a character-limit was exceeded.
    ConnectorAPI key
  • Add a labeled list of offerings (structured snippet) shown beneath a Google Ads ad — a category header plus 3-10 short items, e.g. header "Brands" with values "Nest", "Nexus", "Chromebook", or header "Services" with "Hosting", "DNS", "CDN". Helps searchers self-qualify before clicking. ALSO KNOWN AS: structured snippet, snippet, category list, offering list, header values, product list, service list KEYWORDS: structured snippet, snippet, header, values, brands, services, products, courses, destinations, list, category, google ads, extension WHEN TO USE: - "Add a structured snippet listing our service tiers" - "Create a snippet under 'Brands' with Nest, Nexus, Chromebook" - "Register a list of supported integrations as a snippet extension" - "Show our product categories below our Google Ads" WHEN NOT TO USE: - Want a single phrase (e.g. "Free Shipping") → use create_google_callout_extension - Want clickable links to subpages → use create_google_sitelink_extension - Want a pricing table (with amounts) → use create_google_price_extension - Same snippet for Microsoft Ads / Bing → use create_microsoft_structured_snippet_extension INPUTS: - header: snippet category title (1-25 chars). The Google-Ads vocabulary requires picking one of Google's predefined header names ("Brands", "Models", "Service catalog", "Insurance coverage", etc.) — pass the chosen header string here. - values: 3-10 items, each 1-25 chars. EXAMPLE: create_google_structured_snippet_extension( header="Brands", values=["Nest", "Nexus", "Chromebook"], )
    ConnectorAPI key
  • Reject a mission — it was not a real buying signal. Use this when you want the system to LEARN from the rejection. Each reason maps to a different RL penalty applied to the source feed's scoring weight, so the swarm gets smarter about that product's leads over time. When to use this vs `delete_mission`: - `reject_mission`: the lead was a bad signal — feed it back so the scoring weight adjusts. Always prefer this when you have any opinion on why the lead was wrong. - `delete_mission`: queue cleanup only (duplicates, stale leads, things you don't want to teach the system about). NO learning. DON'T use `reject_mission` when the post is gone (deleted by the poster, 404'd, removed by mods, or stale beyond your reply window). The lead wasn't bad — it just disappeared. Use `delete_mission` instead, which leaves the source station's `rl_weight` untouched. Reaching for `reject_mission(not_relevant)` here penalises a station that did nothing wrong. Valid `rejection_reason` values: - `spam` — bot/promoted/automated post (heaviest penalty) - `not_relevant` — wrong audience or topic - `wrong_product` — relevant signal but wrong product matched it - `too_vague` — signal too weak to act on - `sarcasm` — venting/ironic, not a real buyer - `already_customer` — they already bought (no penalty) - `no_reason` — default if you don't have a strong opinion Pick the most accurate reason — they apply different penalty sizes, so accuracy directly improves how the system learns.
    ConnectorAPI key
  • Mark a relation as no longer true by setting its end date. The relation stays in history (query_world_model with aspect="historical_state") and no longer appears in the current state. Fails if the relation is already closed or if valid_until is earlier than its start date. For works_for, reports_to and located_in changing to a new target, relate_entities with supersede_previous=True closes the previous one instead. Args: relation_id: Relation id (from query_world_model or get_entity_relations). valid_until: ISO 8601 date from which it stopped being true. evidence_ref: Required source for the change (URL, person, document, or episode id).
    Connector
    Destructive
    API key
  • Return pairs of entities that may be the same and that NEXUS's own judge could not decide. They are yours to resolve, not the user's: the user should not be asked about them. Do it at a natural pause, never in the middle of a task. Each doubt carries ``evidencia`` (both entities' type, aliases, relations and facts: what the judge saw) and ``juez`` (its verdict and reason). ``juez.veredicto == "fusionar"`` with ``frenado_por`` means the judge already said they are the same and a safety guard held it back: you would be confirming, not deciding alone. Add what you know about the user and, if a fact is missing, retrieve_memories. Decide each with decide_merge_proposal(por_usuario=False, razon=...): "fusionar" for the same real-world referent (spelling variant, translation, nickname shown in the evidence); "distintas" for related but different things (this one is permanent); "posponer" when you cannot tell. Read-only. Args: limit: How many doubts to return (1-20). Returns {dudas: [{id, score, entity_a, entity_b, juez, evidencia, veces_pospuesta}], pendientes}.
    ConnectorAPI key
  • Close a goal with a terminal status, ending its pursuit. Records the closing event and reason in the goal's history and saves a version snapshot of the goal. A closed goal cannot be reactivated; to stop work temporarily use pause_goal instead. Args: goal_id: UUID of the goal. final_status: One of 'achieved', 'failed', 'expired', 'superseded'. reason: Explanation for closing. Required when final_status is 'failed'; optional otherwise.
    Connector
    Destructive
    API key
  • Ask the hypothesis engine to create one new hypothesis from a knowledge gap it detects in the world model. Hypothesis generation is currently disabled on this server. The call returns HTTP 409 with error 'hypothesis_generation_disabled' and creates nothing. This response is expected and is not a failure. When generation is enabled, the call stores one new hypothesis, or returns 404 if no gap is detected. To read existing hypotheses, use list_hypotheses.
    ConnectorAPI key
  • Turn a mod's display name or part of its id into the exact mod id the other lab tools take, and see which mods the Modkeel lab has data on. Use it first when you only have a name ("Entity Culling", "Sodium") or are unsure of the id; skip it when you already have the id. To download a mod, call find_mod_download with its name instead: it finds the mod itself. Matching ignores case, spaces and punctuation; best match first, then the mods with more lab data. An empty query lists every mod the lab tested. Returns each mod's id, the Minecraft versions it was tested on and how many runs and pairs it has, plus the total number of matches. No match means the lab has no data on that mod, not that it is broken (get_mod_info still finds it on Modrinth by slug). mc keeps only mods with a run on that exact version; limit caps the list, and total says how many matched.
    ConnectorNo auth
  • What one Minecraft mod is and where to get it: its Modrinth page (title, summary, categories, loaders, the Minecraft releases its author lists, client/server side, licence, downloads, source and issue links) next to a short Modkeel lab record (tested versions, run and pair counts). Use it for "what does this mod do?", "is it client-only?", "where do I download it?" or to link a suspect from diagnose_crash. The versions here are what the author declares; for what the lab saw run or crash, use get_mod_runs, and for two mods together, check_mod_pair. Takes a lab mod id or a Modrinth slug; from a name, call search_mods first. The project is matched by the SHA-1 of a jar the lab ran (source: jar_hash, exact), else by slug (source: slug); a mod on neither returns found: false, not an error. Asks Modrinth live on each call (nothing cached here), so it can fail when Modrinth is down: retry later.
    ConnectorNo auth
  • The exact file to install for one mod on one Minecraft version and loader, with its download link. Use it for "where do I get X for 26.2 Fabric?", after diagnose_crash names a mod that needs another build, or when updating a pack to a new version. Checks, best first: the author's own build on Modrinth (status official: file, required mods, and lab runs of that exact file), a community port the Modkeel lab verified in one of its modpacks (port), the author's file for an earlier version that the lab ran on this one and saw start (older_build), a port the lab built from its public source and saw start but does not host (build_from_source: the repo, branch and how to build it), the loaders the author does build for that version (other_loader), else none; when that older file crashed in the lab instead, older_build says so and next_step warns against it. next_step says what the user can do in each case. For what the mod is, use get_mod_info; for whether it ran in the lab, get_mod_runs. Takes the mod's name as the user wrote it ("YUNG's Better Mineshafts"), a lab mod id or a Modrinth slug: no search_mods call needed first. Asks Modrinth live (two to five calls, nothing cached), so it can fail when Modrinth is down: retry.
    ConnectorNo auth