Skip to main content
Security-first AI agent runtime

Build AI agents that can be trusted to act.

Manasvi is a policy-governed runtime for building agents with approval-gated actions, auditable execution, tool mediation, sandboxing, and memory provenance.

agent-runtimeLIVE
01:01intent.receivedactive
02:02policy.evaluatedpass
03:03tool.access.requestedgranted
04:04approval.requiredpending
05:05execution.completedok
06:06audit.event.storedrecorded

One message. Six governed steps.

Manasvi separates conversation, policy, approval, memory, tools, and execution so agents can be useful without becoming blindly autonomous.

Intent
Policy
Tools
Approval
Execution
Audit
Intent — user message received and parsed into an execution intent
Policy — every proposed action evaluated against operator-defined rules
Approval — sensitive or high-risk actions paused for human sign-off
Audit — every outcome written to an append-only, integrity-checked trail

Governance built into every layer

Agents that propose, policies that decide, humans that approve. Every capability is mediated — not bolted on after the fact.

Policy-first execution

Every sensitive action is evaluated against explicit policy before execution. The model proposes; policy decides.

Approval-gated tools

Agents propose actions. Humans or policies approve them. Executors only run approved, cryptographically-signed intents.

📋
Auditable by design

Every decision, tool call, approval, denial, and execution result is written to an append-only, integrity-checked trail.

🛡
Sandboxed runtime

Tool execution is isolated with controlled filesystem, network, secret, and process access. Plugins can't escalate trust.

🧠
Trust-aware memory

Memory is separated by provenance and trust level to reduce poisoning risks and prevent unsafe context reuse across sessions.

🔌
Built for real integrations

Telegram, Gmail, Calendar, filesystem, and web tools can all be governed consistently through the same policy engine.

The model doesn't call tools directly.

In most agent frameworks, the model outputs a tool call and the system runs it. Manasvi puts a governance layer in between. The model proposes. Policy decides. A signed intent is created. Execution is sandboxed. Everything is recorded.

This means you get real control — not just logging after the fact, but actual gates that can stop, redirect, or require approval for any action.

Policy-firstSigned intentsApproval flowsSandboxed executionAppend-only audit
Why this design matters →
Model output
Policy evaluation
Approval gate
Signed intent
Sandboxed execution
Audit record

A governed runtime for trustworthy AI agents.

Free, open source, and runs entirely on your machine. Policy-first execution from the ground up.

Start the quickstart →Learn more