Security

Your account, your keys, your record.

Vocion runs in your own AWS account or on your own host, under the controls your security team already runs. Records stay in your Postgres. Model calls go to the providers you choose. A self-hosted install has no Vocion control plane and does not phone home.

Deployment

AWS first. Or your own host with Postgres 16.

The reference deployment provisions the network, a host, Secrets Manager and DNS in your account with OpenTofu, then runs Vocion with Postgres, Langfuse and TLS on that host. Schedules and workflows run inside the app on its own Postgres. Agent turns can also run on Amazon Bedrock AgentCore Runtime.

Your AWS account or your own host

VOCIONMPL-2.0

app · agents · schedules and workflows · the record

Postgres 16

pgvector · records, documents, durable jobs

Langfuseoptional

traces

OTel collectoroptional

metrics and logs

Caddy

TLS

On AWS: VPC · IAM · Secrets Manager

  • prompts, documents an agent reads, synced documents for search

    Model providers you choose

    Op
    • Chat: Anthropic, OpenAI or Amazon Bedrock, per agent
    • Search: OpenAI or Bedrock embeddings
    • Rerank: Claude Haiku when an Anthropic key is set
  • the repo being built

    Software Factory builds

    • Claude Code · Anthropic
  • what you sync and send

    Tools and channels you connect

    SlTw
    • HubSpot, Gmail, GitHub, Slack, Resend, Twilio
  • traces

    Langfuse Cloud, if you choose it

    • Instead of the Langfuse on your host
Inside the line: what the reference deployment runs. Outside: the services you choose that process data.

Posture

The questions your security team will ask.

Short answers first. Each one comes from how the product works today, and the docs carry the detail.

Where secrets live

Connector credentials are encrypted with AES-256-GCM under a per-workspace key. Wrap that key with your own AWS KMS key, and grant the host role use of it. Runtime config sits in Secrets Manager and is written to the host at boot.

Where data is processed

Prompts and the documents an agent reads go to the model provider you choose per agent: Anthropic, OpenAI or Amazon Bedrock. Every synced document is embedded for search, by OpenAI by default or by Bedrock. Search reranking uses Claude Haiku when an Anthropic key is set. Software Factory builds run Claude Code, so the repo goes to Anthropic. Slack, Resend and Twilio carry the text of messages on those channels.

Where traces go

Langfuse runs beside the app by default. Point it at Langfuse Cloud or turn it off.

Who holds your data

Records live in your own Postgres. Connected systems keep their own copies. Operators with database or Langfuse access can read records, including assistant threads.

Who approved it

Every approval, edit and rejection is recorded with the person and the time. Automatic runs record the rule that let them run.

What it may do

Rules in trust.yaml and the Autonomy page set which actions wait for a person. A fixed list always does.

How replies are checked

Slack, email and SMS webhooks are signature-checked, and a reply decides only a card waiting in that thread, or about a record the thread follows. With several pending, Vocion matches your words to the newest one it names, and confirms which it decided. Email replies are matched by sender address and rely on your domain's DMARC enforcement, so use the app or Slack for protected actions.

How to stop it

A running mission or workflow can be cancelled from the app.

What changed

Agents, teams and rules are YAML in git. Each workspace apply is versioned, and builds, missions and workflows record the version that ran.

What you can audit

The source. Vocion is open source under MPL-2.0, so your team can read every line.

Control

Autonomy is set per action, and recorded.

An action either waits for a person or runs on its own above its confidence bar. Small changes that can be undone can run on their own by default, each listed with Undo. Some plugin rules run on their own too, such as merging a pipeline fix. If someone rejects or undoes work an action did on its own, it steps down a level and is flagged for review. Sending email and connecting a new data source always wait for a person.

Waits for a person

A person approves each run.

Post a chat message

An admin lets it run on its own

Rejected or undone after it ran on its own? It goes back to waiting for a person and is flagged for review.

Small changes that can be undone start here

Runs on its own

Runs above its confidence bar. Below it, waits.

Update a HubSpot recordUndo

Always asks

Built in. No rule or promotion changes it.

Send an emailAlways asks
Illustration.

Set per action, in each workspace.

The Autonomy page lists each action with its setting. The CRM update runs with Undo, the chat post waits for a person, and sending an email from your mailbox always asks.

Illustration. Sample data.

Certifications

On certifications, plainly.

SOC 2 and ISO 27001 certify the operator of a hosted service. Vocion is software you run, so the controls that govern your deployment are your own, which is where your auditors want them. If Metacto runs Vocion for you, ask us for our controls and access policy. What Vocion adds is the record an audit asks for.

Every run recorded

The steps, the tools called and who approved it. Each chat turn, mission and build records its cost.

Every approval named

Customer-facing work links to the person who approved it.

Every change versioned

Each workspace apply is versioned, and changes made in the app are recorded in Vocion.

Get started

Walk your security team through it.

We will run Vocion in a sandbox account, show where every credential, trace and record lives, and answer the questionnaire with the source open.

For your team: a demo built around your job, on sample data.

For your engineers

Run it yourself

The requirements, the AWS path and the repo.