Skip to main content

Your First Workflow

This page walks through what actually happens when you send Manasvi a message — so you understand what the system is doing, not just that it works.

Start everything first

Make sure all services are running. See Run Locally or Quickstart if you haven't done that yet.

Start the terminal chat​

pnpm cli

You'll see a prompt like:

Manasvi terminal  (session: session:abc123)
Type a message, or /help for commands.

You:

Send a simple message​

You: What can you help me with?

The agent will respond with its capabilities. Behind the scenes, here's what happened:

  1. Ingress received your message, verified the source, and normalized it into an internal format
  2. Orchestrator resolved your identity, retrieved any existing session context, and sent the message to the model
  3. The model generated a response — just conversational text, no tool calls needed
  4. Policy evaluated the interaction (even a simple reply is recorded)
  5. Audit recorded the full event

That all happened in the time it took to get a response.

Ask the agent to use a tool​

You: Search for recent news about AI safety

Now the full pipeline activates:

  1. The model decides it needs the web search tool to answer properly
  2. It outputs a tool call proposal — a structured request like "search for: AI safety news"
  3. The policy engine evaluates: is this tool call allowed for this user?
  4. If allowed, an execution intent is created and cryptographically signed
  5. The execution manager verifies the signature before running anything
  6. The tool executes in a sandboxed environment with declared constraints
  7. The result flows back to the model, which formulates a response
  8. Audit records every step: the proposal, the policy decision, the intent, the execution, the outcome

Why this matters: The model never calls the tool directly. It proposes, and a separate system decides whether and how to execute. This separation is what makes governance possible.

Try a multi-turn conversation​

You: That's interesting. What else do you know about AI governance?
Agent: [responds with context from the previous message]

You: Can you summarize the key points in a list?
Agent: [uses previous context to give a structured answer]

Manasvi maintains session context across turns. The orchestrator retrieves relevant memory before each model call, so the agent remembers what you were discussing.

What gets recorded​

Everything you just did was recorded in the audit trail. To see it:

curl http://localhost:4107/audit/events?limit=20

You'll see entries for:

  • Messages received by ingress
  • Policy decisions (even the routine "allow" ones)
  • Tool invocations (if any)
  • Agent responses

The audit trail is append-only. Events can't be modified or deleted after they're written.

Next steps​

Now that you've seen the pipeline working:

  • Connect Telegram — get mobile chat with your agent → Connect Telegram

  • Connect Ollama — run a local model for privacy and no API costs → Connect Ollama

  • Connect Claude — run with Anthropic Claude as your cloud model provider → Connect Claude

  • Understand the concepts — learn what each part does in plain language → Core concepts

  • See the architecture — understand how the services fit together → Architecture overview