SPEAKMAN.AI · Release v1.2.0

A human review gate that isn't a rubber stamp. A revise tool that doesn't start over. A design pass before anyone writes code.

By David Speakman ·

SPEAKMAN.AI is the free, local, open source engine behind the same MCP workflows the hosted platform runs. Version 1.2.0 adds a real human-in-the-loop revision gate alongside the existing automated validator, a delegate mode that hands a step to your own connected coding agent instead of asking you to paste data in by hand, a tool that revises a finished session without re-running it from scratch, and a new workflow that turns architecture and requirements into reviewable screen mockups.

speakmanai.log — real output, this release INFO Update cascade started for MCP_BUSINESS_ANALYST_V1; affected steps: [1, 2, 3, 4, 5] INFO Agent MCP_BUSINESS_ANALYST_V1 requires human review by MCP_BA_HITL_V1. Pausing (loop 0)... INFO Session transitioned COMPLETED -> DRAFT -> COMPLETED. Every downstream agent revised in place.
1
new human review gate type, alongside the existing automated validator.
1
tool to revise a finished session: no full re-run.
1
new built-in workflow. UX & Interface Design.
5
duplicate planner agents found live and collapsed into one.
Who Reviews What

An automated check and a human check are different tools now.

Every agent can declare either, both, or neither. When both are set, the automated validator always runs first on every fresh output. A human's time never goes toward catching what a score already could have.

AI_VALIDATOR
Score-driven
Checks completeness, traceability, schema adherence. Retries automatically below a score threshold.
HITL_VALIDATOR New
Human-driven
Approve, or send back with feedback. For judgment calls a rubric can't make.
Combined
Both, on one agent
Automated gate first, human gate only ever sees output that already passed it.
Why This Release

A pause-for-input step and a real review gate look similar. They aren't.

Pausing a workflow to ask a human for data it can't derive on its own was already there. What was missing was the other direction: a human looking at something the engine already produced and either signing off or sending it back with real feedback, then watching the same agent revise in place rather than start over. That gap showed up constantly once workflows started producing anything a person actually needed to weigh in on rather than just supply.

Revising a finished session had the same problem from a different angle. A requirement changes after the architecture is already done, and the only option was re-running the whole thing, losing whatever decision trail existed and regenerating parts that never needed to change. update_session exists because that's a bad tradeoff every time it comes up.

What Shipped

Four changes. All load bearing.

Review

HITL_VALIDATOR: a real human review-and-revise gate, structurally parallel to the automated validator but scoped to judgment calls a score can't make.

Execution

Delegate-first steps hand real work (fetching a URL, reading a file, running a script, or calling your own MCP server) to your connected coding agent instead of pausing for a human to do it by hand.

Revision

update_session reopens a completed session and revises one agent's output. The change cascades through everything downstream, and nothing unaffected gets touched.

Workflow

A new built-in pipeline. Architecture and requirements go in; a screen inventory, flow diagrams, a proposed design direction, and clickable mockups come out. Detailed below.

The New Workflow

From architecture JSON to a mockup someone can actually click through.

Seeded from a completed Solution Architecture Document and, where available, its requirements contracts, MCP_UX_DESIGN_V1 maps logical components to real UI surfaces and requirements to screens, checks that nothing traces back to a gap, then proposes a design direction and builds self-contained HTML mockups for the highest-priority screens. The mockup step runs through your connected coding agent rather than a fixed background model. Design judgment is exactly where the model behind that connection matters most.

SAD + Requirements Screen Inventory Coverage Check Flow Diagrams Design Direction + Mockups

This is the same workflow the /generate-ux-design skill calls into as Phase 2.5 of the SPEAKMAN.AI SDLC pipeline, shown below.

Where This Sits

One engine. Free or hosted, it runs the same pipeline.

The hosted SPEAKMAN.AI platform runs a full idea-to-production SDLC pipeline. This repository is the free, local, MIT licensed engine that executes the same MCP workflows, with full functionality for an individual user and no account required.

Idea Intake

/generate-project-brief

A raw idea, coached into a structured brief. No SPEAKMAN.AI session, no credits.

Architecture

/generate-sad

Business description in, Solution Architecture Document out.

Requirements

/generate-requirements

Domain model, naming dictionary, use cases, API and DB schema.

UX Design

/generate-ux-design This release

Screen inventory, flow diagrams, a proposed direction, and clickable mockups.

Code Generation

/generate-speakmanai-code

A working, milestone gated, git committed codebase.

Infrastructure

/generate-infrastructure

Terraform for GCP, AWS, and Azure, plus pre-pentest hardening.

Pentest & UAT

/pentest

A gated security engagement, run against the live staging build.

Verified Live, Not Just Merged

We found five copies of the same agent and traced why one of them was broken.

Every workflow needs exactly one planner agent, identified by type rather than a hardcoded ID. A live database check found five separate copies doing that job instead of one shared definition, including a planner with a completely empty prompt, silently live, because a fallback that should have caught it only triggers on a missing key, not an empty one.

Before planner_agent.get("systemPrompt", DEFAULT) → "" (empty string used as-is)
After planner_agent.get("systemPrompt") or DEFAULT → real prompt, every time
Runs On Your Machine

The free tier is not a trial. It is the whole engine.

ModeCommandDatabase
Windows exeRun SpeakmanAI.exe~/.speakmanai/speakmanai.db
Docker + SQLitedocker-compose up --buildspeakmanai_data volume
Docker + MongoDBdocker-compose --profile mongo upMongoDB, port 27017
Dev, no Dockeruvicorn server:app --port 8000~/.speakmanai/speakmanai.db

There's no account and no cloud dependency. No per-call fee beyond your own provider key, either. Workflows are plain JSON on disk, readable and editable rather than locked inside a hosted black box.

Individual · Free · MIT

Run the engine locally.

Download the exe, or clone and run with Docker. You get full functionality on your own API keys, with nothing sent anywhere else.

Teams · Hosted

Need multi-tenant support and audit trails? Metered billing too?

The hosted SPEAKMAN.AI platform runs the same engine with the infrastructure teams need on top of it.

Written by David Speakman. Speakman Consulting designs and builds this kind of system for growing organizations: agent workflows with the governance that keeps humans in the loop.