LM-Kit OneThe Private AI Application Server
Promoted
Launching today
ContextsBase - Backlog for Coding Agents
From chat prompts to autonomous backlog execution.
45 followers
From chat prompts to autonomous backlog execution.
45 followers
ContextsBase is a single-source context infrastructure platform (web app) that manages unified project knowledge for AI agents. Served seamlessly over MCP, it holds your specs, rules, data models, workflows, and guidelines in one place. It isn't limited to web development. Whether you're building backend APIs, mobile apps, systems software, or even managing marketing campaigns and brand engines, it acts as the central source of truth that your agents read from and report back to.
ContextsBase - Backlog for Coding Agents
ContextsBase - Backlog for Coding Agents
What inspired this:
I was shipping a side project almost entirely with Claude Code and Cursor. The code was good. The product was wrong. The agent invented a tier field on Customer that didn't exist and every new session started with me pasting the same Notion doc back into the chat. When a teammate changed a rule on Thursday, the code from Tuesday was quietly out of date and nothing told us.
Every one of those failures was the same failure. The agent never had the actual product truth, and nothing checked its work against it.
The problem I was trying to solve
Every team has a knowledge base for humans: a wiki, a data model diagram, acceptance criteria. Agents get a pasted prompt. ContextsBase is the knowledge base your coding agents read from. You write features, business rules, the data model, and test cases once. Your agent, whichever one you use, pulls them over MCP in one call, builds the feature, writes tests that assert the business rule instead of clicking a button, and records what it built against which spec.
We don't supply the agent. Bring Claude Code, Cursor, Copilot, or Windsurf. We supply what they build from.
How it evolved
The first version was just structured specs. Then I noticed the test problem: everybody's E2E suite proves the UI works, almost nobody asserts the rule. So test cases became Given/When/Then business rules that the agent turns into Playwright code. Then came iterations, an ordered queue agents pull from one step at a time so two agents never build the same thing. Recently we added a theme designer, so agents build on-brand from published tokens, and a page editor where you click any element on your live site, change the words or styling, and your agent applies it to the real code.
AskDeck
"a spec instead of a pasted prompt" is the whole thing, nice framing. Question from a non-coding use: our deck agent has the same problem with people, not repos. A caller's brand, their usual audience, the template they always want. Is ContextsBase strictly code-shaped (features, data model, tests) or could it hold that kind of context served over MCP too?
ContextsBase - Backlog for Coding Agents
@thicreator "A spec instead of a pasted prompt" is a universal pattern for any complex AI agent. Whether an agent is writing React components or generating 15-slide pitch decks, serving structured, versioned specs over MCP ensures the agent always acts with exact, up-to-date domain context without relying on users to prompt it correctly every time.