Toolbelt gives AI agents a shared, queryable workspace for documents, structured data, entities, relationships, and persistent findings.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install toolbelt
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install toolbelt using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is Toolbelt?
Toolbelt is a collaborative data substrate exposed through a single MCP server. Agents can upload approved documents, automatically extract entities and relationships, and query structured tables, unstructured documents, vector indexes, knowledge graphs, SQL data, and timelines without manually stitching multiple systems together.
This Openclaw Skills integration supports persistent, cross-session collaboration. Agents can record relevant findings, propose reusable lessons, share namespace access, and let future agents build on curated context. It is designed for tasks that benefit from hybrid retrieval, automatic knowledge extraction, or shared state—not simple one-off lookups.
Toolbelt Use Cases
- Ingest approved documents and make extracted entities and relationships immediately queryable.
- Ask natural-language questions spanning SQL tables, documents, and knowledge-graph relationships.
- Maintain persistent findings that remain available across agent sessions and different MCP clients.
- Coordinate multiple agents or teammates around a shared Toolbelt namespace.
- Search documents semantically with vector retrieval when a focused similarity search is needed.
- Run precise SQL queries over registered structured tables.
- Inspect entity profiles, traverse relationships with Cypher, or review chronological timeline events.
- Propose governed business definitions, ownership metadata, corrections, and reusable agent guidance.
- Use Openclaw Skills when a task requires a shared data workspace rather than isolated agent context.
How Toolbelt Works
- Detect whether the MCP client is already connected by calling
toolbelt_list_namespaces. - If no valid connection exists, ask for explicit consent before sending the anonymous onboarding request to
https://app.toolbelt.ai/api/onboard. - Capture the returned bearer token, MCP URL, user ID, namespace ID, and expiration timestamp. The default anonymous account provides 1,000 calls, one namespace, and a 30-day lifetime.
- Ask for separate consent before writing the Toolbelt MCP entry to the user's client configuration. The entry points to
https://mcp.toolbelt.ai/mcpand uses the bearer token in an Authorization header. - Call
toolbelt_context(namespace_id)to inspect available assets, suggested next actions, and server-provided routing guidance. - Hand the task to the appropriate MCP tool: use
toolbelt_askfor general hybrid questions, or use vector, SQL, entity, graph, record, timeline, knowledge, save, register, and sharing tools according to the task. - Return a concise connection status including the MCP URL, namespace, account tier, and expiration without exposing the full token.
- If persistence beyond 30 days or higher free quotas are needed, ask for an email and complete the optional claim and verification flow.
- For Pro or Team upgrades, direct the user to the Toolbelt web UI; agents do not process billing or credit cards.
Toolbelt Setup
Requirements
Use an MCP-compatible agent such as Claude Code, Claude Desktop, OpenClaw, Cursor, Windsurf, Gemini CLI, or Codex CLI. No Toolbelt account is required initially because the skill can provision an anonymous account after user approval.
Provision an account
The agent must obtain explicit consent before making the onboarding request. The approved request is:
curl -X POST https://app.toolbelt.ai/api/onboard \
-H "Content-Type: application/json" \
-d '{}'
Store the returned token, mcpUrl, user.id, namespace.id, and expiresAt only as needed for MCP configuration. Treat the token as a credential.
Configure the MCP client
After separate user confirmation, configure the MCP server. For Claude Code:
claude mcp add toolbelt \
--transport http https://mcp.toolbelt.ai/mcp \
--header "Authorization: Bearer <token>"
For JSON-based clients, add this entry to the client's MCP configuration:
{
"mcpServers": {
"toolbelt": {
"url": "https://mcp.toolbelt.ai/mcp",
"headers": {
"Authorization": "Bearer <token>"
}
}
}
}
Common configuration paths include ~/.openclaw/mcp.json, ~/.cursor/mcp.json, ~/.codeium/windsurf/mcp_config.json, ~/.gemini/mcp.json, ~/.codex/mcp.json, and Claude Desktop's platform-specific configuration file. Reload or restart the MCP client if required.
Security and account lifecycle
Never write configuration files without consent, echo the full bearer token, or upload unapproved files. Remove the toolbelt MCP entry or revoke the token through https://app.toolbelt.ai to disconnect access. Anonymous accounts expire after 30 days; email verification upgrades an account to the Verified tier.
Toolbelt Data Schema & Taxonomy
Workspace model
| Object | Purpose | Key metadata |
|---|---|---|
| Account | Identifies the Toolbelt user and entitlement tier | user.id, tier, quota, expiresAt |
| Namespace | Shared workspace boundary for agent data | namespace.id, name, accessible assets |
| Document asset | Uploaded unstructured content | Source file, extracted entities, relationships, searchable text |
| Structured table | Queryable tabular data | Table registration, columns, SQL-accessible records |
| Entity | Knowledge-graph profile extracted from data | Entity identity, attributes, linked records |
| Relationship | Edge connecting entities or records | Relationship type, source context, graph traversal data |
| Timeline finding | Persistent agent-recorded insight | Finding content, relevance, chronological placement |
| Lesson draft | Proposed reusable agent guidance | Rule or lesson text, approval status |
| Governed knowledge fact | Curated business definition or ownership rule | Term, definition, owner, bindings to tables and documents, approval state |
| Share credential | Namespace access URL for another agent or teammate | Namespace binding, access capability, revocation state |
Retrieval and persistence surfaces
toolbelt_askorchestrates hybrid retrieval across SQL, documents, and the knowledge graph.toolbelt_vectorsperforms semantic vector search over documents.toolbelt_sqlruns exact queries over structured tables.toolbelt_entityandtoolbelt_graphread profiles and graph relationships.toolbelt_recordwrites relevant findings to the persistent timeline.toolbelt_savepersists an asset, whiletoolbelt_register_tableexposes an existing Kinetica table without re-ingestion.toolbelt_timelinereads chronological findings andtoolbelt_contextreturns workspace orientation and routing guidance.
Account tiers
| Tier | Calls per month | Storage | Namespaces | Lifetime or access |
|---|---|---|---|---|
| Anonymous | 1,000 | — | 1 | Free, 30 days |
| Verified | 2,000 | 1 GB | 10 | Free after email verification |
| Pro | 150,000 | 50 GB | 50 | $29/month through the web UI |
| Team | 500,000 | 100 GB | Unlimited | $89/month through the web UI |
Toolbelt Advanced Features
- Hybrid retrieval routes a single natural-language question across semantic search, SQL, and knowledge-graph data.
- Automatic knowledge extraction converts uploaded documents into queryable entities and relationships.
- Persistent timelines let later agents retrieve findings recorded by earlier agents.
- Multi-agent collaboration uses namespace sharing so agents and teammates can work from the same state.
toolbelt_lesson_proposesupports governed, reusable guidance that requires workspace-owner approval.toolbelt_knowledgeexposes approved business terms, definitions, ownership, and data bindings, with Mermaid output available for diagrams.toolbelt_knowledge_proposeenables reviewable corrections to extracted edges and proposed governed facts.- Existing Kinetica tables can be registered without re-ingesting their data.
- The MCP server provides per-tool routing guidance through its tool descriptions.
- Consent gates protect network requests, filesystem writes, uploads, and namespace sharing.
- Token handling limits exposure: credentials remain in the MCP client configuration, are never fully echoed, and can be revoked by removing the configuration entry or using the web UI.
- Openclaw Skills users can connect the same Toolbelt namespace across supported MCP-compatible agents and sessions.
SKILL.md
Loading
Related Openclaw Skills
Battle Report Briefing
Turn historical battles and military conflicts into rigorously researched, single-screen mobile PNG briefings with tactical analysis and source citations.
arXiv Decision Intelligence Report
An evidence-first Openclaw Skills workflow that transforms arXiv literature into validated field intelligence, technical options, maturity assessments, and actionable decisions.
AI Tech Insight Compass
AI Tech Insight Compass turns daily AI trend research into source-backed, publication-ready technical articles.
COPPA Guard
COPPA Guard is a local, pre-release scanner that detects children’s privacy compliance risks in product copy and returns prioritized remediation guidance.
McKinsey Library Consultant
A source-grounded consulting delivery workflow that transforms scattered client inputs into validated storylines, executive presentations, strategy documents, and speaker scripts.
Implement
An implementation workflow that isolates code changes, validates them, commits and pushes the result, and opens a draft GitHub pull request.
Featured*
