The engine behind your agents.

SPEAKMAN.AI is a workflow engine built for the agentic era. Plug it into Claude Code, the Claude app or Microsoft Copilot Studio and run multi-agent workflows with human review gates on your own machine. Your assistant can do every AI step itself, so you don't need a model API key until you want background runs.

Not another app to manage.

Open source on GitHub → Start a pilot

Run locally with the Windows exe. No Python, no Docker, no account required.

● ONLINE

A workflow engine your agents can call directly.

SPEAKMAN.AI is not a UI you live in. It's an MCP-native workflow engine your agents connect to, execute against, and consume results from without leaving the tools driving your business.

Not a chatbot.

Not another platform to manage.

A structured workflow engine your pipelines consume directly.

THE PROBLEM

AI tools that demand you live in their UI

Most AI platforms are built around their interface, not yours. Every time your workflow needs intelligence, you switch context, prompt manually, copy output, and paste it somewhere else.

Context switching between tools

Manual copy-paste from AI outputs

No human sign-off on AI-generated decisions

Custom integrations that break with every update

The cost is not a single invoice. It's the compounding overhead of every team living in the wrong tool.

THE POSITION

Works where you already work.

SPEAKMAN.AI exposes multi-agent workflows over MCP. Your agents call it. Your pipelines consume it. Your team gets structured, reviewable intelligence delivered directly into the tools they use.

Claude Code calls a workflow. Gets a complete output.

Copilot Studio runs a Solution Architecture workflow. The document comes back to your agent.

Your pipeline triggers a workflow. Structured data comes back.

The result? Multi-agent intelligence embedded in your existing tools. Zero new UI to manage.

Your agents call it. Not the other way around.

OPEN SOURCE — LOCAL VERSION

Run it locally. Own it completely.

Individual users get everything from the open source local engine (MIT licensed), with no subscription. No account and no cloud dependency. A fresh install needs no model API key either: your connected assistant handles the AI steps, and you can add your own key whenever you want the engine to run them in the background. It runs in Claude Code and the Claude app, and it has been tested end to end in Microsoft Copilot Studio.

Windows Exe

Download and run. System tray icon, browser setup page. The 17 skills ship inside the exe, and one click installs them into Claude Code. No Python or Docker installation required.

Docker + SQLite

Single command. Persistent data. Workflows seeded automatically on first launch.

Docker + MongoDB

Same features as SQLite mode just swap the database if you prefer MongoDB or already run it in your stack.

Run From Source

For contributors. Start the server with uvicorn in a Python environment, no Docker needed.

Provider	Notes
Gemini	Recommended - gemini-3.5-flash-lite / gemini-3.8-flash / gemini-3.1-pro-preview
Vertex AI	The same Gemini 3.x models, through your Google Cloud project
Claude	claude-sonnet-5-5 / claude-opus-5-5
OpenAI	gpt-6-luna / gpt-6.1-sol / gpt-6-astra

Download on GitHub →

AVAILABLE WORKFLOWS

Start with what's built. Extend from there.

Sizing Estimate

A quick gate before any architecture work. Breaks a raw idea into components and gives each a relative S/M/L/XL size, never a dollar figure, with a human review step. ~5–10 minutes.

Solution Architecture

Full multi-agent pipeline. Six specialist agents (business context clarifier, business analyst, application architect, technical architect, compliance officer and visualization specialist) build a complete SAD, with validators and human review gates between stages. ~10–20 minutes.

Capability Generator

Converts a plain-text description of your tech stack and processes into a structured capabilities JSON array. ~2–5 minutes.

Requirements Engineering

Turns an architecture, or a business description alone, into implementation-ready contracts: domain model, use cases, naming dictionary and a full OpenAPI schema. ~10–15 minutes.

UX Design

Maps requirements to a screen inventory, proposes a design direction and builds clickable HTML mockups for review before any code is written. ~5–10 minutes.

Your workflow

Transform your existing business processes into AI-ready pipelines. Describe your steps and the Workflow Builder generates a ready-to-import SPEAKMAN.AI workflow with agents, HITL pause points, and all. We can help design and build those too.

HOW IT WORKS

One MCP connection. Full workflow access.

Add the SPEAKMAN.AI MCP server to Claude Code, Copilot Studio, the Claude app or your own pipeline.
Call list_workflows to see available workflows.
Call start_session — the workflow runs asynchronously.
Poll with poll_workflow and respond to input requests as they arrive.
Call get_outputs and get_output to consume the structured results.

Your agent does the work. You review the output.

Claude Code — terminal

claude mcp add --transport http --scope user speakmanai http://localhost:8000/mcp

# Or press the Claude Code button on the setup page.
# It runs this for you and installs the skills too.

CONTROL AND TRUST

Structured. Reviewable. Owned by your team.

Every workflow run is scoped to a tenant and linked to a user. Nothing disappears into a chat history.

Reviewable

Editable

Owned by your team

The engine handles execution. People handle judgment.

COMPARISON

Why teams connect SPEAKMAN.AI instead of building their own

Build it yourself

Months of infra work. Multi-tenancy, auth, billing, and orchestration before a single workflow ships.

Generic AI platforms

Live in their UI. Copy output manually. No human sign-off. No integration with the tools you already use.

SPEAKMAN.AI

MCP connection. Instant workflow access. Open source local version, or hosted for teams when you need it.

CLOUD

Built for teams

Multi-tenant

Every workflow is scoped to a tenant. Isolation and access control built in.

Credentialed

API key auth, per-agent provisioning, and concurrency limits on every request.

Your model account

Every model call goes to the provider you choose, on your own account.

Agentic HITL

Workflows pause for human input and resume from the UI or from your agent.

What typically takes months to build is available on day one.

TEAMS

For teams that need more than one machine.

The local engine is for one person and needs no subscription. When a team needs to share workflows and outputs, SPEAKMAN.AI runs in the cloud, still on your own model account.

Two ways to run it

Team edition

Your own tenant on shared infrastructure. Set up when a team needs it.

Private tenant

A dedicated deployment in a region you choose, such as Canada.

Both are priced per seat and set up on request.

What's included

MCP access, tested with Claude Code, the Claude app and Microsoft Copilot Studio
All workflow types, including human review gates
Shared workflows and outputs, with access scoped by tenant and role
Personal and shared item library
Document downloads in Markdown and Word
Every model call on your own account: you choose the provider and model for each tier
Email support

When it makes sense

Several people sharing workflows and outputs.

Data that has to be stored in a particular region.

A workflow that has outgrown one person's machine.

Start with a pilot. The cloud is where your team keeps it running.

Start a pilot

The engine behind your agents.

Multi-agent intelligence. MCP-native. Open source locally. Ready for teams when you need it.

Get it on GitHub Talk to us

Not another app to manage.