The fabric that connects AI to your organization's tools, data, and rules — woven per user, per team, per company
No-code wizard: Import an OpenAPI spec → auto-generates a plugin with tools, connectors, and security settings.
Code option: Write Python, create a manifest.json defining tools, resources, prompts, and security requirements. Package as a ZIP.
Manifest includes: Tool definitions (name, description, parameters), required connectors, minimum role, security toggles, knowledge sources.
Upload: Drag a ZIP into the Visual Designer or upload via POST /api/plugins/upload
Auto-registration: The MCP engine reads the manifest, registers all tools with FastMCP, creates connector entries, and sets up knowledge indexes.
2-click install: Upload → Enable. That's it. No server restarts needed.
Connectors: Define which external services the plugin talks to. Each connector has credential fields (API keys, tokens, OAuth).
Credentials: Set at user, group, or org level. The vault cascade resolves the right key at runtime.
Security toggles: Enable/disable write operations, require credentials, set rate limits, enable audit logging — all per plugin.
Knowledge sources: Add documents, URLs, or text to the plugin's RAG index. Scoped by user/group/org.
Flip the power switch → the plugin's tools become available in the agent loop. The LLM can now discover and call them.
Disable: Flip it off → tools instantly disappear from the LLM's toolset. No prompt changes, no code changes. The security wrapper blocks all calls.
Hot-reload: No pod restart needed. Enable/disable is immediate.
Publish: Package your plugin and list it in the marketplace with description, pricing, and security audit report.
Discover: Browse plugins by industry, capability, or rating. One-click install.
Revenue share: Plugin developers earn 70% of sales. ContextWeaver takes 30% platform fee.
Certified: Plugins undergo security audit before listing. Certified badge = meets ContextWeaver security standards.
When the AI needs context to answer a question, it searches across 5 priority levels. The most specific knowledge wins:
P5 Personal: Your private notes, preferences, saved items. Only you can see these.
P4 Group: Your team's shared knowledge — budgets, protocols, team docs. Visible to team members only.
P3 Organization: Company-wide policies, handbooks, vendor lists. Everyone in the org sees these.
P2 Plugin: Knowledge specific to a plugin — API documentation, tool instructions.
P1 Engine: Global defaults — currency, language, audit policies. Broadest scope.
Deduplication: If the same topic appears at multiple levels, higher priority wins. Personal preferences always override company defaults.
Each piece of knowledge is stored in a scoped index with ACLs. The search layer filters results by the requesting user's identity and group membership.
Index naming: {engine}-{plugin}-{scope}-index
ACL enforcement: Even if someone guesses an index name, the search API checks _user_email and _user_groups before returning results.
Connectors define HOW a plugin talks to an external service. Each connector specifies credential fields, OAuth flows, and endpoint URLs.
Per-user credentials: Each user can configure their own API keys for a connector. The vault cascade finds the right key at runtime.
OAuth support: Google, Microsoft, GitHub OAuth flows built in. Users click "Connect with Google" and their token is stored in the vault.
Extensible: Import any OpenAPI spec → auto-generate a connector with all required fields.
When a tool needs an API key, the vault checks three levels in order:
1️⃣ User key — Does this specific user have their own key? (most secure)
2️⃣ Group key — Does their team/department have a shared key?
3️⃣ Org key — Is there a company-wide fallback key?
If no key found at any level → tool call is blocked. The AI reports "credentials required" and suggests the user configure their key in settings.
Storage: Azure Key Vault (production), HashiCorp Vault, or local encrypted file. The AI never sees raw keys.
Every toggle is enforced in the MCP engine's security wrapper — at the code level, not in prompts. No amount of clever prompting can bypass them.
Per-plugin: Enable/disable, require credentials, allow/block write operations, rate limiting, audit logging.
Per-engine: Enable/disable entire engine, block cross-engine calls, require MFA for admin, maintenance mode, audit all calls.
Visual Designer: Toggle switches are visible right on the canvas. Click 🔐 to toggle auth, 🔒 for read-only, 💰 for approval required.