Tvix MCP
Explore the MCP tools, resources, and prompts that power Tvix agent integration. Learn how agents store, search, and retrieve Memory and Wiki context through the Model Context Protocol.
Tvix connects to AI agents via the Model Context Protocol (MCP), providing a standard way for agents to store, search, and retrieve Memory and Wiki context.
At a Glance
| Type | Name | What it does | Who uses it |
|---|---|---|---|
| Tool | add_memory | Save personal context to long-term memory | Agent calls automatically when it learns something worth remembering |
| Tool | search_memory | Search past memories by meaning | Agent calls when past context would help its response |
| Tool | get_current_date | Get current date and time in a timezone | Agent calls before date-filtered memory search |
| Tool | add_wiki | Add a factual knowledge document to the wiki | Agent calls when user shares stable, reference-style knowledge |
| Tool | search_wiki | Hybrid search across wiki documents | Agent calls when factual knowledge would improve its response |
| Tool | update_wiki | Edit an existing wiki document | Agent calls to keep a document current |
| Tool | delete_wiki | Permanently remove a wiki document | Agent calls with user confirmation |
| Resource | tvix://profile | Read-only user settings profile | Agent reads when stable profile context matters |
| Resource | tvix://recent | Read-only recent memory timeline | Agent reads for latest updates and recency checks |
| Prompt | start | Guidance for memory workflow and safety | Agent can load at session start |
You don't need to call any of these yourself. Your agent handles everything automatically based on the conversation.
Memory vs Wiki
Tvix exposes two complementary stores through MCP:
- Memory (
add_memory,search_memory): Personal context that describes the user: preferences, decisions, habits, ongoing projects. - Wiki (
add_wiki,search_wiki,update_wiki,delete_wiki): Factual knowledge that the user wants to keep as reference material: docs, specs, notes, and anything else that is stable and addressable by title.
For user questions, agents should generally call both search_memory and search_wiki, then combine the results. See the Knowledge Wiki guide for details.
MCP Server
Tvix runs a remote MCP server accessible at:
https://mcp.membase.so/mcp
| Property | Value |
|---|---|
| Transport | Streamable HTTP |
| Authentication | OAuth (browser-based) |
| Package | tvix (npm) |
For MCP setups, the npx -y tvix@latest --client <client> CLI command handles registration and auth automatically for supported clients. See the Agents guide for per-client setup.
Note: Claude Code users should start with the Tvix Claude Code plugin. The plugin exposes the same Tvix memory and wiki tools directly inside Claude Code, plus slash commands, hooks, and project-aware workflows. The remote MCP server remains available for other clients and as a Claude Code fallback.
Tools
Tvix exposes tools that agents can call during conversations. They split into two groups: memory tools for personal context and wiki tools for factual knowledge.
Memory Tools
Stores long-term memory that persists across sessions. Agents use this **proactively** when the user shares durable context worth remembering: preferences, recurring habits, ongoing projects, goals, constraints, etc. Transient one-off information is skipped unless the user explicitly asks to save it.
| Parameter | Type | Required | Description |
| ---------- | ------------------------- | -------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `content` | string (max 50,000 chars) | Yes | The memory content to store |
| `project` | string (max 60 chars) | No | Project or category slug to file this memory under. Agents should only set this when the user **explicitly** mentions a project name, tag, label, or category (e.g. "save this under project X", "tag this as Y"). Do not guess or infer. See [Projects](#projects). |
| `metadata` | object | No | Leave empty for normal agent usage. |
Searches stored memories by **semantic similarity**. Agents use this when the user asks to recall something from a previous session, or proactively when past context would improve their response. Results are returned as episode-centric bundles with related context.
| Parameter | Type | Required | Description |
| ----------- | ------------------------ | -------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `query` | string (max 1,000 chars) | Yes | Natural-language search query. Use empty string to fetch recent memories. |
| `limit` | integer (1-30) | No | Max results to return (default: 20) |
| `offset` | integer | No | Pagination offset (default: 0) |
| `date_from` | ISO date/datetime string | No | Inclusive lower bound for date filtering |
| `date_to` | ISO date/datetime string | No | Inclusive upper bound for date filtering |
| `timezone` | IANA timezone string | No | Date-only parsing timezone, for example `America/Los_Angeles` |
| `sources` | array of strings | No | Filter by source. Supports integrations (Slack, Gmail, Google Calendar), AI clients (Cursor, Claude, Claude Code, VS Code, ChatGPT, Codex, Gemini CLI, OpenCode, Poke, OpenClaw, Hermes), imports (ChatGPT, Claude, Gemini), and dashboard-created memories. If omitted, all sources are searched. |
| `project` | string (max 60 chars) | No | Restrict search to a single project slug (exact match). Use only when the user explicitly asks for a project or category scope. See [Projects](#projects). |
Results can include a **relevance score** (0–1) when available, indicating how closely each match fits the query.
Returns the current date and time in the user's Tvix timezone (and UTC). Agents call this before `search_memory` to resolve relative phrases like "today", "yesterday", "this week", or "next weekend" into ISO 8601 `date_from`/`date_to` values.
This tool takes no input parameters. The response contains four fields:
| Field | Description |
| ------------ | ------------------------------------------------------------------------- |
| `now_utc` | Current UTC timestamp in ISO 8601 (e.g., `2025-01-15T02:30:00.000Z`) |
| `timezone` | The effective IANA timezone used (from the user's saved Tvix timezone) |
| `local_date` | Local date in that timezone (e.g., `2025-01-15`) |
| `local_time` | Local time in that timezone (e.g., `11:30:00`) |
Projects
Memories can be grouped into projects (lightweight category/tag slugs). Both add_memory and search_memory accept an optional project field.
- User-driven, not inferred: Agents should set
projectonadd_memoryonly when the user explicitly mentions a project, tag, label, or category (for example, "save this underacme-rewrite", "tag this aspersonal"). Agents must not guess or invent a project. - Known projects hint: The Tvix MCP server injects the current user's existing project slugs into each tool's description, so agents can reuse existing slugs instead of creating new ones for the same concept.
- Filtered search: Pass
project: "acme-rewrite"tosearch_memoryto restrict results to that project. Exact slug match only. - Managed in the dashboard: Users can move memories between projects and filter by project from the Memories tab. See the Memory guide.
Tip: Users who want hands-off project tagging can drop a line in their profile instructions such as "When I mention
acmeorside-project, tag those memories under the corresponding project." The agent will then apply the tag automatically during matching conversations.
Wiki Tools
Creates a new document in the user's knowledge wiki. Agents use this for **factual, reference-style knowledge** (documentation, specs, stable notes). Personal preferences and habits should go through `add_memory` instead.
| Parameter | Type | Required | Description |
| --------- | -------------------------- | -------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `title` | string (1-500 chars) | Yes | Title of the wiki document |
| `content` | string (max 100,000 chars) | Yes | Complete markdown content. Preserve reports, discussions, analyses, tables, examples, and decisions unless the user explicitly asks to save a summary. |
| `project` | string (max 200 chars) | No | Project name to file this document under. New Projects are created on first use. Leave empty when the user does not name a Project; Tvix routes safely or falls back to Basic. |
Document creation respects the user's [plan quota](/features/wiki#document-quotas). If the limit is reached the call returns a quota error.
When the tool succeeds, agents should tell the user the returned destination when present, such as `Saved to Project: Docs` or `Saved to Basic`.
Searches the user's wiki using **hybrid search**: a full-text keyword index (BM25-style) and semantic similarity, fused with Reciprocal Rank Fusion (RRF). Results include each document's full body so the agent has enough context to answer immediately.
| Parameter | Type | Required | Description |
| ------------ | ------------------------ | -------- | ----------------------------------------------------------------------------------------------------- |
| `query` | string (max 1,000 chars) | Yes | Natural-language or keyword search query. Use empty string `""` to fetch recent wiki documents. |
| `limit` | integer (1-20) | No | Max results to return (default: 10) |
| `project` | string (max 200 chars) | No | Restrict results to a specific Wiki Project by name. If omitted, all Projects and Basic are searched. |
| `collection` | string (max 200 chars) | No | Legacy alias for `project`. Prefer `project` for new clients. |
Edits an existing wiki document. Use `search_wiki` first to resolve the document ID.
| Parameter | Type | Required | Description |
| ------------ | -------------------------------- | -------- | ----------------------------------------------------------------------------------------------------------------------------- |
| `doc_id` | UUID | Yes | ID of the document to update |
| `title` | string (1-500 chars) | No | New title |
| `content` | string (max 100,000 chars) | No | Complete replacement markdown content. Preserve the full updated artifact unless the user explicitly asked for summarization. |
| `project` | string or `null` (max 200 chars) | No | Move the document to a Project by name. New Projects are created on first use. Set `project` to `null` to move it to Basic. |
| `collection` | string or `null` (max 200 chars) | No | Legacy alias for `project`. Prefer `project` for new clients. Set to `null` to move to Basic. |
When the result includes a destination such as `Moved to Project: Docs`, `Moved to Basic`, or `Current destination: Basic`, agents should report it to the user.
Permanently deletes a wiki document. This action cannot be undone; agents should ask for user confirmation before calling it.
| Parameter | Type | Required | Description |
| --------- | ---- | -------- | ---------------------------- |
| `doc_id` | UUID | Yes | ID of the document to delete |
Resources
MCP resources provide read-only data that agents can pull into their context:
| Resource | URI | Description |
|---|---|---|
| User Profile | tvix://profile | Read-only MCP resource that mirrors Settings > Profile (display_name, role, interests, instructions, timezone) |
| Recent Memories | tvix://recent | Top 10 recent memories, ordered by event time when available and capture time otherwise |
Prompts
MCP prompts inject pre-built instructions and context into the agent's system prompt:
| Prompt | Description |
|---|---|
| start | Teaches the agent when to use search_memory, when to use resources, when to save memory, and what to avoid storing. |
The start prompt can be used at session start to provide:
- Workflow guidance for memory retrieval and memory write behavior
- Safety guidance for secrets and transient information
- Resource usage guidance for profile versus recent timeline reads
Recommended Agent Behavior
Use these patterns for best results:
- For questions about saved knowledge, past conversations, documents, decisions, or user-specific context, search Memory and Wiki together when both could matter. Use Memory for personal context and Wiki for factual/reference knowledge.
- For stable user settings, agents can read the MCP resource
tvix://profile. It mirrors the dashboard Profile settings: display name, role, interests, custom instructions, and timezone. - For latest memory context or recent-memory timeline questions, agents can read
tvix://recent. This is a Memory timeline, not a Wiki or integration changelog. - For date-filtered memory searches, call
get_current_datewhen needed, then pass explicit ISO date filters tosearch_memory. - Save durable user context with
add_memory; skip secrets and one-off transient data unless explicitly requested. - Save factual, reference-style knowledge with
add_wiki. Preserve the full artifact unless the user explicitly asks for a summary. If the user names a Project, pass it inproject; otherwise leave it empty and report the returned destination.
Connection Setup
For detailed setup instructions for each agent, see the Agents guide.
Cursor
One-click install
Claude Code
Plugin recommended
Claude
Custom connector
ChatGPT
Manual config
Next Steps
Memory
Explore and manage memories.
Knowledge Wiki
Deep dive on Wiki documents, Projects, imports, and search.
Chat in Dashboard
Talk to your knowledge base without an external agent.