Mentat — multi-agent development harness
A harness that runs an entire software development workflow through four coordinated AI
agents — Orchestrator, Architect, Developer, and Staff — instead of a single chat
session. A human-approved ADR is a hard gate before any code is written, and the handoff
from Architect to Developer travels as a validated handoff.json.
Wired to ClickUp, GitHub, Slite, and Figma over MCP — originally for Claude Code, now also running in opencode — with a self-hosted Prometheus + Grafana stack that tracks token cost and usage per agent session.
Docs: devrkd.github.io/mentat →
Source: github.com/devrkd/mentat →
Morpheus — unified LLM inference service
A Bedrock-style unified inference service: one HTTP API, one model id per request, any
provider. Ask for claude-opus-5 or gpt-5 and Morpheus routes
the call to the matching vendor — Anthropic or OpenAI — translating the request and the
response both ways. Conversation transcripts live on the server, so a session can switch
providers mid-conversation and the new model still sees the whole history.
A small Python service with a built-in chat UI, API-key authentication with per-team model allowlists and spend ceilings, permission-gated tools (built-in and MCP), streaming, and a full test suite that runs with no credentials and no network.
Docs: devrkd.github.io/morpheus →
Source: github.com/devrkd/morpheus →
Fluent — AI writing assistant
Fluent fixes spelling and grammar, or rewrites text for clarity and tone — but only when
you ask. Type a trigger like /fixit or /rewrite after your text,
and the suggestion appears beside the field; nothing changes until you accept it.
The same idea is built three ways: a native macOS menu-bar app in Swift (AppKit), a Chrome and Firefox browser extension, and a standalone Node/TypeScript service — each calling Anthropic Claude and OpenAI GPT.
The repository is called rewrite; the project is named Fluent.