Joltrin
Joltrin AGENTS

Verification barrier for AI agents

VERIFICATION BARRIER FOR AI AGENTS

The agent does the work. Joltrin checks it first.

An agent can say a backup is validated when it never ran one. Joltrin keeps its own record of the steps that ran and refuses the next step until the steps it depends on are in that record. It does not use what the agent claims. Durable memory and recovery sit underneath, so a restarted run can pick up where it stopped.

Agent Memory

Agent state, step checkpoints and vector similarity search that persist across sessions, in the same process as the agent.

Agent Recovery

Resume an interrupted run from its last checkpoint. A transaction that does not commit rolls back.

Agent Verification

Each step is checked against the recorded trace and the runbook's rules before it commits. A blocked step returns what is missing.

Agent Governance

Paid editions add team workspaces and a hash-chained audit log. Enterprise adds Okta or Microsoft Entra ID sign-in.

WORKED EXAMPLE

An agent migrating a database

An agent is asked to migrate a database. Joltrin stores its checkpoints, checks each sensitive step against the steps already recorded, and refuses a step whose preconditions are missing.

Without Joltrin No barrier
Agent ──────► Tool / API ──────► Irreversible Action
  • An agent that says the backup is done can still drop the database
  • A crash loses what the agent was holding in memory
  • A failed step can leave partial changes behind
  • Nothing records which steps really ran
With Joltrin With the barrier
Agent ──► Memory ──► Verification ──► Tool ──► Safe Action
  • The drop is refused until a backup is validated in the trace
  • A restarted run resumes from its last checkpoint
  • A transaction that does not commit rolls back
  • The trace records which steps ran

Where it fits

Runbooks you could write

AI DevOps & Cloud Automation

Require a reviewed plan before an apply, and a validated backup before a schema drop.

Software Engineering Agents

Require the tests step before a merge step, with checkpoints to resume from after a failure.

Enterprise Workflow Agents

Require an approval step before a payment or a customer data update runs.

Multi-Agent Swarms

A shared transactional store lets several agents coordinate state without racing each other.

Tool-Using AI Agents

Allowlist the tools each agent may call, and refuse a claimed reading that no tool returned.

MCP and A2A

Register the barrier with an MCP client using sop-mcp-server setup, or call it over A2A.

SAME ENGINE BEHIND THE REAL MCP + A2A SERVERS

Stop an agent before it drops your database.

This runs verify, the safety and reachability barrier compiled to WebAssembly. It is the identical Go package that gates tools/mcpserver and tools/a2aagent, Joltrin's real MCP and A2A servers, just wired to buttons here instead of protocol calls. Try dropping the database before validating a backup.

Runbook: db-maintenance

try step 3 first

Four missions

0 of 4 done

Each one is checked against the real barrier's answer, not against anything pretend. Clearing them is saved in this browser only.

    Barrier log

    // waiting for WASM engine...

    A block is feedback, not a dead end. The result: line is what an agent receives: the rule that tripped (blocked_by), the state that is missing (missing_state), and the steps that would establish it (established_by_steps). Try the steps it names, in order. See how real agents used it.

    What this page is, and isn't

    This is a static site with no backend, so it cannot run a live MCP or A2A network server, both are real protocols an agent connects to over stdio or HTTP. What you just clicked through is the actual verification logic those servers call before committing a step. Run the real servers yourself:

    git clone https://github.com/SharedCode/joltrin.git
    cd joltrin
    go run ./examples/verify_barrier        # same scenario, in a terminal
    go run ./cmd/sop-mcp-server              # MCP over stdio
    go run ./cmd/sop-a2a-agent &             # A2A over HTTP
    curl localhost:8087/.well-known/agent-card.json

    NEXT

    The barrier checks the steps. Private Intent checks who said yes.

    Write your own rule, have other systems sign the proof and one exact approval, and get a receipt the tool holding the keys must see before it acts. Try to cheat it in your browser. Nothing is sent anywhere.

    Try Private Intent
    Next step

    Want this on your own agents?

    The barrier is open source and stays free. Pro is for teams that need one shared policy for every agent and a record of each allow and block. Enterprise adds Okta or Entra ID sign-in and custom policy rules. Tell us what you are building and we will follow up.

    More details (optional)

    No credit card needed. We reply by email.