Hi team I would consider myself a Graphite and MCP...
# how-do-i
m
Hi team I would consider myself a Graphite and MCP power user separately but I'm still wrapping my head around using these two together. Could you share what you think a successful Graphite MCP workflow looks like? Are you saying for example Cursor generates a large complex plan, you tell it to break it up into small diffs and commit along the way with the Graphite MCP? That sounds like the dream but generally coding agents don't nail the approach on the first try and there are often lint, typescript and test failures. I don't love the idea of making a large stack of diffs with random errors in each diff. Let me know if I'm thinking about this incorrectly 🙂
p
(If applicable) To help us more clearly understand what you're trying to do, please provide additional details about: - Screenshots / debug logs / list of commands you've tried to run so far - Your workflow - Current state
Hey Mike! Have you seen this great blog post by my colleague Kush?
m
Thanks for sharing this brett, this is helpful but more or less how I would envision it working myself. I would love to hear more detail on what this would look like doing a feature end to end and verifying each diff is mergable. Example: • Cursor plan mode, - build me log in feature. make sure to do small stacked diffs ◦ creates a great plan with diffs • Execute plan • The agent writes solid code, moves on to the next diff ◦ the problem when I try this is I'll often have lint, and typecheck errors at this stage and can't commit ◦ Then I would have a stack of 6 diffs that all have random linting and type errors that need fixing and QA Curious if you guys have a way to optimize this. Thanks!
p
Hey Mike, just to close the loop here.. i had a chat with the author of that blog post and they suggested
the mcp includes some instructions to verify each incremental state before continuing though we've also found luck internally by reinforcing this with local context (ie. claude.md) i'm sure they can recreate with cursor. asking the agent to "verify each commit builds and passes lint" is usually enough
hope this helps!