v1.3.0

Headline features

1. render_document — server-side document rendering, no local Python or Node required
A new MCP tool converts a Markdown document into a fully self-contained, branded HTML document entirely on the server side — no local script, no Node/puppeteer dependency, nothing the calling client needs installed. This is what unblocks running generate-sad and generate-compliance-report from any MCP client, including ones that can't shell out to a local interpreter at all.

2. Confirmed working: Microsoft Copilot Studio
generate-sad has been run end-to-end inside a live Copilot Studio agent, connected to a local SPEAKMAN.AI server over a tunneled connection. SPEAKMAN.AI's MCP surface was already client-agnostic in principle; this is the first outside-Anthropic client it's been verified against.

Under the hood

• PDF generation and the local template-directory system (templates/default/, templates/corporate/) are both retired — HTML is the deliverable now, print-to-PDF-able from any browser, with no per-user template setup required.
• The default document branding is a plain text wordmark rather than an embedded logo image — a full rendered document now runs about 1,400 tokens, down from what an unoptimized embedded logo would have cost (~57,000 tokens for the same document).
• Fixed a long-standing bug where a workflow step listing more than one agent in its agents array silently ran only the first one — the planner had always been free to bundle multiple agents into a single step, but the engine only ever executed one. Steps are now split and executed individually regardless of how the planner grouped them.
• Fixed a bug where reaching the HITL review-loop cap discarded the human's final round of feedback and advanced using stale, pre-feedback content instead of the freshly regenerated output.
• RE_USE_CASE_ANALYST_V1 previously produced exactly one use case per functional requirement by construction; it now enumerates every use case a requirement actually implies (access patterns, CRUD completeness, screen-level states, actor variations), roughly 5x the catalog density on a real test project.
• RE_DOMAIN_MODELER_V1 and RE_API_SCHEMA_DESIGNER_V1, previously single large agents, are each decomposed into focused sub-agents plus a mechanical synthesis step — reducing single-call cognitive load and the blast radius of any one regeneration.

v1.2.1 — Expanded HITL review coverage

Every built-in workflow now has a human sign-off gate on the agents whose output ships straight into a delivered artifact with nothing else checking it first.

Requirements Engineering gained six HITL_VALIDATOR gates — use case catalog, domain model, field definitions, naming dictionary, API schema, and database schema — replacing two older plain-pause review steps with the same revise-in-place loop the rest of the platform already uses.

Three more workflows picked up a gate apiece: the Compliance Officer in Solution Architecture (previously the only agent in that pipeline with zero checks of any kind), the UX Needs Analyst in UX Design (the screen-inventory step everything downstream is built on), and the Sizing Context Clarifier in Sizing Estimate (the only agent reading raw business input, with nothing downstream able to catch a wrong reading).

Also fixed: a dangling dependency typo in CLAUDE_WORKFLOW_CREATOR_V1 left over from before the shared planner was consolidated under one name.

Full details: RELEASE_NOTES.md

v1.2.0

What's new

• Human-in-the-loop revision gates (HITL_VALIDATOR) — a real human review-and-revise gate on any agent, structurally parallel to the existing automated AI_VALIDATOR but driven by a person. When an agent has both, the automated check always runs first.
• Delegate-first execution (executionMode: "delegate") — any step can hand off to your connected coding agent instead of pausing for a human to supply data by hand. New companion generator: generate-speakmanai-skill-workflow.
• update_session — a new MCP tool that reopens a completed session, revises one agent's output, and cascades the change through everything downstream in place, guarded by the first genuinely atomic state transition in this codebase.
• New workflow: UX & Interface Design (MCP_UX_DESIGN_V1) — screen inventory, requirement-traceability check, a proposed design direction, and self-contained HTML mockups for human review, before any code gets written.
• Consolidated five drifting copies of the workflow planner agent into one shared definition — found live, including one with a silently empty prompt that a fallback bug let slip through
• Fixed a dedup bug where a human review loop could leave stale, superseded sections in an assembled project document
• Automated validators now actually read the dependency context their own prompts assumed they had
• Every workflow-generator meta-tool now understands HITL_VALIDATOR as a first-class option

Full announcement: https://consulting.speakman.ai/blogs/speakmanai-v1-2-0

See README.md / RELEASE_NOTES.md for full configuration details.

v1.1.0

What's new

• Standard model tier — a third abstract tier (fast / standard / advanced) sitting between the existing two, configurable via STANDARD_MODEL
• Vertex AI support in the setup UI — pick "Vertex AI" as a provider and configure GCP project/region directly (uses Application Default Credentials, no API key needed)
• Workflow database seeding fix — new bundled workflow files are now backfilled into an existing database on startup instead of only seeding once on a completely empty database
• Fixed several model-tier resolution bugs (Anthropic tier collapse, fast-tier override, stale fallback defaults)
• gemini-3.1-flash-lite-preview → gemini-3.1-flash-lite (Google promoted it to general availability; the old preview alias no longer resolves)
• Desktop exe no longer inherits Docker-oriented .env settings (fixes a Vertex AI credentials-path bug specific to the exe build)

Full announcement: https://consulting.speakman.ai/blogs/speakmanai-v1-1-0

See README.md / SETUP.md for full configuration details.

v1.0.0 — Initial Release

SPEAKMAN.AI Local — v1.0.0

First public release of the SPEAKMAN.AI local multi-agent workflow engine.

What's included

• Windows desktop exe — no Python or Docker required
• Four built-in workflows: Solution Architecture, Capability Generator, Workflow Creator (standard + Claude-optimised)
• Multi-provider LLM support: Gemini, Claude, OpenAI, Ollama
• Browser-based setup page with one-click MCP client configuration
• SQLite database with automatic workflow seeding on first launch
• Docker deployment option for teams (SQLite or MongoDB)

Getting started

1. Download SpeakmanAI.exe below
2. Run it — a system tray icon appears and your browser opens automatically
3. Enter your API key on the setup page
4. Connect your AI assistant and run your first workflow

See the README and SETUP.md for full documentation.

Requirements

• Windows 10/11 (64-bit)
• An API key for your chosen LLM provider (Gemini, Claude, or OpenAI) — or a local Ollama installation
• An MCP-compatible AI assistant (Claude Code, Claude Desktop, or Cursor)