Skip to content
SYNAPLAN

AGENTS // SYS-03

AI agents with RAG, tools and human approval.

  • AI agents
  • RAG
  • Human approval
  • MCP tools
  • Ollama
  • Desktop skills
  • Audit log

A Synaplan AI agent receives a message from WhatsApp, Outlook or your website chatbot, retrieves from your RAG knowledge base, reasons with Ollama or a cloud model, calls tools and MCP servers, waits for human approval, and executes. Every step is in the audit log. Self-hosted.

Mission 0427T+00:00
inbound
outlook/mail
agent
procurement-assistant
knowledge
/contracts /suppliers
model
llama-local
tools
erp.search mail.send
node
de-dus-k8s-03
Approval required
  1. message received
A demo — but every line maps to a real capability.
00 /Mission profile
  1. 01 understand

    Classify intent, language and channel.

  2. 02 retrieve

    RAG over your documents and memories.

  3. 03 reason

    Local or cloud model — your routing.

  4. 04 tool

    ERP, calendar, mail, MCP servers.

  5. 05 approval

    A person confirms before it writes.

  6. 06 execute

    Do the thing. Log everything.

01 /Knowledge first

Agents read your documents before they open their beak.

Retrieval-augmented generation over uploaded files, connected drives and mailboxes. Per-user memories add what the person prefers. Sources are cited.

  • 01

    Documents

    PDF, Office, mail, images with OCR — chunked and embedded.

  • 02

    Memories

    What the user told you last month, stored per user, visible and deletable.

  • 03

    Scopes

    Which agent may read which folder is a setting, not a hope.

02 /Tools and skills

Payloads: what the agent may actually do.

Tools are explicit: built-in actions, plugins, MCP servers, and Agent Skills that run on the user's own computer through the desktop client.

  • 01

    Built-in

    Web search, document generation, calendar, mail, file operations.

  • 02

    MCP

    Any MCP-compatible tool server. Synaplan speaks the protocol both ways.

  • 03

    Desktop skills

    Runs locally under a folder allowlist. No shell is ever constructed.

03 /Mission control

A human approves before anything is written.

Approval gates are part of the pipeline. Pending missions show up where people already work — in Synaplan, in Outlook, on the phone. Approve, reject, or edit and approve.

  • 01

    Approvals

    Per tool, per agent, per amount — policy decides what needs a person.

  • 02

    Telemetry

    Every retrieval, model call and tool execution in the audit log.

  • 03

    Stop button

    Turn an agent off from the same row. The consequence is stated in one sentence.

T-00:00:10

Give your AI a job.

Build a first agent on our platform in minutes, or read how the pipeline works in the docs.

[ Start on the platform ]Agent docs ↗