Skip to main content

Architecture Overview

Manasvi is built as a set of cooperating services, each with a narrow responsibility. No single component has unchecked authority — decisions are made by the right service, verified by the next one in the chain.

System diagram

Ingress Plane
┌──────────────────────────────────────┐
│ Telegram │ Slack │ API Gateway │
└──────────────────┬───────────────────┘
│ normalized message

Orchestration Plane
┌──────────────────────────────────────┐
│ Agent Runtime │
│ (planning loop, proposal parsing, │
│ prompt injection detection) │
│ │ │
│ Policy Service │
│ (allow / allow_with_approval / deny)│
│ │ │
│ Approval Flow │
│ (optional human sign-off) │
└──────────────────┬───────────────────┘
│ signed execution intent

Execution Plane
┌──────────────────────────────────────┐
│ Execution Manager │
│ (intent verification, dispatch) │
│ │ │
│ Sandbox Runtime │
│ (network + filesystem constraints) │
└──────────────────┬───────────────────┘
│ (optional remote dispatch)

Node Plane
┌──────────────────────────────────────┐
│ Node Manager │ Node Agent(s) │
│ (dispatch signing, nonce tracking) │
└──────────────────────────────────────┘

Supporting Planes
┌──────────────┐ ┌────────────────┐ ┌──────────────────┐
│ Memory Plane │ │ Extension Plane│ │ Audit & Logging │
│ (trust- │ │ (plugins, │ │ (immutable trail│
│ classified │ │ tool registry)│ │ of all actions)│
│ stores) │ │ │ │ │
└──────────────┘ └────────────────┘ └──────────────────┘

Planes

Manasvi's architecture is organized into logical planes — groups of components with related responsibilities:

PlaneResponsibility
IngressAccept messages from external channels, normalize into internal format
OrchestrationThink, plan, authorize — the governance layer
ExecutionRun tools safely within declared constraints
NodeDispatch workloads to remote execution environments
MemoryStore and retrieve context, organized by trust level
ExtensionPlugin lifecycle and tool registry
AuditRecord everything for accountability

Security flow

Every consequential action flows through the same verification chain:

  1. Agent runtime — parses model output into a structured proposal
  2. Policy service — decides allow/deny/approve
  3. Approval flow — if required, waits for human sign-off
  4. Intent issuance — creates a signed, time-limited execution intent
  5. Execution manager — verifies the intent before dispatching
  6. Sandbox — enforces declared constraints during execution
  7. Node agent — re-verifies if dispatched to a remote node

No step trusts the previous step's claim — each verifies independently.

Design principles

  • Separation of concerns — each plane has a narrow, well-defined job
  • Fail-closed — uncertainty defaults to denial, not permission
  • Signed artifacts — authorization is cryptographic, not assumed
  • Auditable — every decision and action is recorded

Built-in tool layer

Manasvi ships with a governed built-in toolset that gives it immediate practical utility:

ToolPurposeRisk
tool.local-file-readRead local files in sandboxLow
tool.http-fetchFetch remote content under egress policyMedium
tool.web-searchWeb search with structured resultsMedium
tool.memory-note-writeWrite notes to governed memoryMedium
tool.approval-requestRoute actions to human reviewersHigh
tool.shell-commandBounded shell execution (approval required)High

Every tool invocation flows through the same security chain: policy → intent signing → sandboxed execution → output validation. Tool results carry explicit trust classification and provenance.

Built-in tools overviewDefault tool sets