The end-to-end journey from login to agentic AI action. Explore how authentication, plugin building, knowledge ingestion, policy cascades, and real-time tool execution come together in one unified platform.
Back to HomeSecure identity resolution from first click to active session
End users land on a streamlined chat interface with engine selection. No admin clutter — just natural language interaction with policy-aware AI.
Route: /chat → engine dropdown → SSE streamAdmins access the complete Cerebro dashboard: plugin builder, visual designer, engine management, monitoring, and user administration.
Route: /dash → role: admin → full navCustom-branded Entra ID login with organization logo, SSO options, and conditional access policies pre-evaluated.
Endpoint: /auth/login → 302 redirectFull OAuth 2.0 + OIDC flow with MFA enforcement, conditional access evaluation, and token issuance with custom claims.
Protocol: OIDC • Grant: auth_code + PKCEHttpOnly secure cookie set with encrypted session. X-User-Email header propagated to all downstream services for identity context.
Token: HttpOnly cookie • Header: X-User-EmailFrom natural language description to production-deployed AI engine
Describe your plugin in natural language. The AI generates a complete plugin structure: plugin.yaml (metadata & config), tools.py (MCP tool handlers), resources.py (MCP resources), and prompts.py (prompt templates). Edit inline, test immediately.
Compose engines visually: attach plugins from the library, wire connectors (GitHub, Jira, Email), add knowledge sources, and set security policies. The canvas auto-validates dependencies and generates the engine manifest.
Canvas → engine.yaml manifest generationAttach knowledge sources at any scope — URLs, documents, wikis, raw text. Azure AI Search indexes with vector + hybrid search. Five-level scope hierarchy: Engine → Plugin → Org → Group → Personal. Each index tagged with scope for policy cascade resolution.
Azure AI Search • Vector + Hybrid • 5-level scopeHelm chart generation from engine manifest. Kubernetes deployment with namespace isolation, Workload Identity binding, secret injection. Plugin distribution from Hub library. Connector registration with health verification. Full rollback on failure.
Helm → K8s deploy → health check → rollbackAll plugins version-controlled in a GitHub repository. Changes tracked, diffs reviewed, rollbacks simple. The plugin library stays in sync with the Git source of truth.
Repo: ajaypunreddy/contextweaver-pluginsThree-pillar architecture powering the ContextWeaver platform
The most important flow: from natural language prompt to agentic AI action
The agent reads MCP resources — cerebro://indexes to inventory all available knowledge sources and cerebro://status to verify backend health. This builds an awareness map of what knowledge and tools are available before any action is taken.
Calls search_index across every relevant index. Results are scored via the 5-level scope cascade: USER_PREFERENCE > GROUP_POLICY > ORG_POLICY > PLUGIN_KNOWLEDGE > ENGINE_KNOWLEDGE. Policy-tagged results are treated as hard limits — the agent cannot override them.
Merges defaults from base indexes with user-level and team-level overrides. Secrets are identified and masked — never shown to the user or included in LLM context. A merged values table is constructed for tool execution.
Merge: base defaults → team overrides → user overridesLoaded plugins expose MCP tools — jira-search-issues, github-create-pr, email-send. Before executing, the agent searches knowledge for applicable policies. [ORG_POLICY] and [GROUP_POLICY] results are hard limits that constrain execution. User preferences are additive guidance.
For action requests, the agent generates resolved CLI commands from templates + merged parameters. Commands execute via background jobs with real-time step-by-step progress streaming back to the user interface.
Background job → SSE progress stream → completionAll search contexts, resolved parameters, and tool results feed into the LLM for streaming RAG synthesis. Tokens stream in real-time to the user interface. The model is constrained by policy-tagged context — hard limits shape the response.
LLM: Azure OpenAI • Streaming: SSE tokensStructured citations link every claim back to its source document, URL, and originating index. Users see exactly where information came from and can verify against the original source material.
Format: [source_title](url) — index: scope_tagFive-level priority system governing every AI decision
Defense in depth across every layer of the platform
Per-user API tokens stored in Azure Key Vault. Each user’s credentials isolated, encrypted at rest with HSM-backed keys. No shared secrets.
Each engine runs in its own Kubernetes namespace with a dedicated service account. Workload Identity for zero-secret pod authentication to Azure services.
Every tool invocation logged with identity, timestamp, parameters, result, and MCP trace events. Full audit trail for compliance and debugging.
Per-tool security controls: read_only, approval_required, min_role. Dangerous operations require explicit human approval before execution.
Secrets are never displayed in UI, logs, or LLM context. Key Vault references render as 🔐 *** in all interfaces. Zero secret leakage.
Terraform modules for Azure, AWS, and GCP. Swap cloud providers via configuration. No vendor lock-in at the infrastructure layer.
Engines expose tools, resources, and prompts over standard MCP transports
Any MCP-compatible client can connect: GitHub Copilot, Claude Desktop, ContextWeaver Chat, and Custom Apps. Standard protocol means zero integration code.
/sse
HTTP /mcp
Dash /dash
list_indexes List all available knowledge indexessearch_index Search a specific index with querysearch_all Search across all accessible indexesask_cerebro Full RAG query with citationscerebro://indexes Index inventory and metadatacerebro://status Backend health and readinesscerebro://plugins Installed plugin manifestrag_query Structured RAG query with policy awarenesscompare_indexes Compare content across indexessummarize_index Summarize an entire indexFrom first login to verified AI-powered action
Entra SSO authenticates the user seamlessly. A clean chat interface opens with an engine dropdown — select your team’s AI agent and start prompting immediately.
The settings panel shows which connectors need credentials (GitHub PAT, Jira API key, etc.). Each credential is stored in Key Vault per user identity — never shared, never exposed.
Type in natural language. The agent performs: policy search → credential resolution → tool execution → streaming result with full pipeline trace. Actions happen in real-time with progress indicators.
Every response includes citations linking to source documents. A pipeline trace panel shows which tools were called, which indexes were searched, and timing for each step. Full transparency.