One MCP server, three agents, zero copy-paste
n8n catches the signal, Specky decides what it means, and Claude Code or Codex writes the change — all against the same MCP server, with a human on exactly one gate.
- Steps
- 5
- Tools
- 6
- Metrics
- 3
- Read
- 4 min
A solo technical founder with a shipped product, an n8n instance, a Claude Code subscription, and no interest in being the message bus between their own tools.
Composite blueprint. This is the loop and the wiring — assembled from workflows our early design partners run — not a single named customer’s audited results. Every number on this page is a metric to instrument, not a result we’re claiming.
What the week looked like
The workflow was: read feedback, open a chat window, paste the feedback, paste the codebase context, paste last week’s decision, hope.
Every agent started from zero. None of them knew what had already been decided or why.
Context pasting was the actual job. The agents were fast; the human bottleneck was clipboard work.
Two agents given the same task on different days produced contradictory changes, because neither could see the other’s reasoning.
Every agent reads from the same Product Graph over MCP. Specky is the shared source of truth — so Claude Code, Codex, and n8n are working from one set of facts instead of three private guesses.
The tools this loop runs on
Native connectors where they exist, and an honest label where they don't — anything else reaches Specky over MCP, the REST API, or an automation hop.
Routes inbound signals — forms, webhooks, inboxes, RSS — into Specky. Can call the MCP server directly as a tool node.
The shared brain. search_workspace, read_workspace_item, create_document, create_tickets, feedback_to_spec, and the rest of the toolset over Streamable HTTP.
Reads the approved spec and its evidence, implements, opens the PR.
Same server, same tools — for the work you’d rather run in the other harness. Both see identical context.
The interactive lane: query the graph mid-edit without leaving the editor.
PRs land with the spec linked. Merges sync back as delivery signals and close the opportunity.
How it actually runs
Every step is labelled by who does it. The steps marked 'you review' are the ones worth keeping human.
n8n stops being glue and starts being a router
AutomaticAny inbound signal — a form response, a webhook, a parsed inbox, a scraped changelog — gets normalised in n8n and pushed into Specky. n8n’s job is transport; it no longer needs to hold any product judgement.
Specky becomes the shared source of truth
AutomaticThe graph holds every signal, decision, PRD, and shipped outcome. Because it’s an MCP server, that Product Graph is readable by any agent you point at it — which is the difference between three agents and one team of agents.
The one human gate
You reviewSpecky drafts the opportunity and the PRD into the PM Inbox. You approve the problem. This is the only step that doesn’t get automated, and it’s the one that makes the rest safe to automate.
Claude Code or Codex picks up the work
AgentPoint either harness at the same MCP endpoint. The agent calls read_workspace_item for the approved PRD, follows its cited evidence, writes the change, and opens a PR. No pasted context, because the context is a tool call.
The merge feeds the graph
AutomaticMerged work syncs back and attaches to the opportunity it came from. Next week’s agent reads that history — so the loop compounds instead of resetting every Monday.
The numbers to watch — not the ones we'd like to quote
We're not going to invent a percentage for you. These are the three metrics that tell you whether the loop is real, and how to compute each one from data the graph already holds.
Context-pasting events per shipped change
Zero
Count the times you manually paste product context into an agent. It’s a crude metric and it’s the honest one — if it isn’t falling, the MCP wiring isn’t being used.
Agreement between harnesses
Claude Code and Codex reach the same conclusion from the same spec
Run both against one approved PRD occasionally. Divergence means the spec is underspecified — which is a spec problem you can fix, not a model problem you can’t.
Approved-to-merged time
Hours
PRD approval timestamp → merge timestamp. Once the human gate is the only synchronous step, this is the number that tells you whether the automation is real.
Start here, not everywhere
Get a workspace API key
Specky → Settings → Developer → Create API key. Keys are scoped to a single workspace, prefixed sk_live_. One tool, feedback_to_spec, is callable with no key at all if you want to test the endpoint first.
Add the MCP server to every harness you use
https://www.specky.space/api/mcp over Streamable HTTP, with Authorization: Bearer sk_live_…. Claude Code, Codex, Cursor, Claude Desktop, and n8n all speak it. Setup snippets for each are on the /for-agents page.
Automate everything except the approval
MCP tools create in draft state by default and land in the PM Inbox. Keep it that way. An agentic loop with no human gate doesn’t ship faster — it ships wrong faster.
https://www.specky.space/api/mcpClaude Code · Codex · Cursor · Claude Desktop · ChatGPT · n8n
Frequently asked questions
Which agents can connect to Specky over MCP?+
Any MCP client. Claude Code, Claude Desktop, the Claude apps, OpenAI Codex, Cursor, Windsurf, Zed, VS Code’s agent mode, and n8n’s MCP node all work — Specky implements the Streamable HTTP transport, so the client list is whatever the ecosystem supports, not a list Specky has to maintain.
Can I use Specky purely as an MCP server, without the web app?+
Yes. The graph, search, document creation, ticket generation, and research campaigns are all reachable as MCP tools. Plenty of technical founders never open the UI except to approve things — and the PM Inbox is the one surface that’s genuinely worth opening.
Do agents get unrestricted access to my workspace?+
No. Every key is scoped to one workspace, every tool call is checked against that scope, and writes default to draft state for human review. Plan-based rate limits apply, and keys are revocable from Settings → Developer.
Run this loop on your own product.
One source, one agent, one approval gate. You’ll know inside a week whether it holds.
14-day free trial · No credit card required · Cancel anytime