| As a daily Claude Code and Cursor user, I kept hitting the same wall:
the agent shows me a plan as a numbered list in the terminal, then
starts coding. I can't reorder the steps. I can't monitor what's
happening at step 8 without scrolling through a wall of output. I
can't attach my Stripe API key to just the payment step without it
leaking into every other step. And I can't tell step 3 to use the
Figma MCP while step 7 uses Supabase. Overture is an open-source MCP server that intercepts the agent's
planning phase and renders it as an interactive node graph in the
browser before execution starts. Each step becomes a node. Dependencies become edges. You click into
a node and attach exactly what it needs: files, docs, API keys, MCP
servers, natural language instructions, all scoped to that step only.
You drag to reorder, delete what you don't need, add what's missing.
At decision points, you pick an approach or run branches in parallel
and compare the output. During execution, each node updates
individually so you see exactly where things stand. Under the hood, agents generate plans in structured XML (<node>,
<edge>, <field> elements). Overture parses it into the graph. Agents
can also declare typed input fields per step, so if a step needs a
Stripe key or a database URL, you fill it in during planning instead
of getting interrupted mid-execution. Works with any MCP-compatible agent. npx overture to try it.
Everything local. Biggest open problem: getting agents to consistently produce
well-structured plans. Claude Code cooperates well. Cursor is
hit or miss. If anyone has experience with prompt engineering for
structured output from coding agents, I'd love to compare notes. GitHub: https://github.com/SixHq/Overture.git |