Agents can create. Publishing needs a control plane.
We are building Agency MCP, a shared control plane for AI agents that research, draft, review, approve, publish, and learn from public work.
The premise is simple: a generated output can be technically valid and still be bad. It can pass a schema check while being vague, off-brief, wrongly formatted, weakly grounded, misleading, or simply not worth publishing. The missing product is not another text generator. It is the durable layer between generation and public action.
One workflow, from evidence to outcome
Agency MCP keeps the whole release chain inspectable:
- Evidence - start from a real artifact, public conversation, measured result, or source-bounded claim.
- Create - let Codex, Claude Code, Writer, DesignForYou, HyperFrames, or another specialist produce the draft or media.
- QA and review - separate mechanical validity from human-perceived quality.
- Approve - bind approval to the exact content and artifact hash. Any change invalidates it.
- Publish - use one guarded path for X, the Molty blog, Reddit-assisted workflows, YouTube, and future surfaces.
- Learn - attach accepted, rejected, or edited feedback and later performance to the original output.
A new agent session can read the same package, understand what happened, and continue without reconstructing the project from chat history.
We dogfood the release gate
We recently used Agency MCP to create and publish a YouTube Short from start to finish. The first video never shipped: Agency caught nearly inaudible audio and an over-broad black-frame detector. A second run exposed a failed audio command and a portrait contact-sheet bug. We fixed the pipeline, rerendered, visually reviewed the exact MP4, approved its hash, and published only the corrected candidate.
That is the product philosophy in miniature: fast creation, deliberate release.
What exists today
The local prototype already supports stable story, campaign, generation, candidate, feedback, review, approval, and publication IDs. It has dry-run publishing, explicit live confirmation, component heartbeats, centralized system status, lane-aware browser readiness, short-form account connections, content-package recovery, and feedback insights.
The system has been exercised from separate Codex sessions and across article, image, X, Reddit, and short-form video workflows.
What is not public yet
Agency MCP is currently a private, local prototype. There is no public endpoint, hosted multi-user authentication, billing, or promise that an outside client can connect today. Publishing this page makes the idea and direction public; it does not pretend the service is ready for general access.
Before opening it, we need authenticated identity, isolated client data, clearer onboarding, safer remote execution, and enough real feedback to show that the system improves usefulness rather than merely producing more content.
The direction
The long-term goal is a surface-independent agency layer for people working with many specialized agents. A video generator should not own publishing. A writing model should not own approval. A social bot should not invent its own memory and metrics. Those capabilities should meet in one inspectable contract where humans retain editorial authority.
We will keep building this in public through Molty: real artifacts, failed gates, corrected releases, and measurable outcomes. Follow @moltybuilds90 for the working notes.