How to make beads work with Cursor AI IDE

[Helleeni]

I successfully installed beads, did init, added mcp server and created a cursor rule 'beads.mdc' that essentially is the content of 'bd onboard' file. 'beads.mdc' is running always at the beginning of each thread.
However, it seems that agents ignore beads - no items are added in the database. Does anyone have a working Cursor - beads installation?
beads.md

[robertocarvajal]

I am attempting the same, seems like aider will not really run commands on it's own, you can use /run inside aider, but as far as hands-off automatic integration that seems to not be something aider is interested to do. I already have a markdown system for issue tracking and I'm curious/exploring beads as a replacement but so far no luck on hooking this up with aider.

[steveyegge]

@robertocarvajal if Aider really can't run tools on its own, it sounds like someone needs to have a talk with the Aider team. That's a serious, serious flaw.

[robertocarvajal]

At this moment I'm not sure if it's intentional or if it's just low in the priority list for the project.. unfortunately that project is really ran by a single dev with no one else helping to maintain it and at the current time he is busy on other projects, so it's been sitting still since at least August, there is an "aider community experiments" fork that adds support for MCP but I really don't like MCP, never did, and now it seems it's been discarded for simpler solutions even by Anthropic..

Anyways, Aider seems to favor a human in the loop approach and giving the LLM limited capacity, for example, by default you have to confirm if you want to add files to the context that might be useful, etc. there is a specific flag to automatically say "yes" but that intentionally disables the or ignores the /run command instructions, so it really seems that Aider philosophy is about having the human ultimately always in control rather than leaning on fully autonomous / agentic workflows at least as of right now.

[steveyegge]

Hi @Helleeni and @robertocarvajal - thanks for raising this discussion!

I wanted to follow up on the Aider integration topic. After reflecting on my initial response, I realize I came across as dismissive of Aider's human-in-the-loop design philosophy. That wasn't my intent, and I apologize for that tone.

Aider Integration Now Supported! 🎉

I've implemented full Aider support for beads. The key insight from @robertocarvajal's comment was spot-on: Aider's design is fundamentally different from autonomous agents like Claude Code, and that's actually a valid architectural choice that many users prefer.

What's New

You can now set up beads for Aider with a single command:

bd setup aider

This creates:

• .aider.conf.yml - Config that loads beads workflow instructions
• .aider/BEADS.md - AI-facing instructions (loaded via read: directive)
• .aider/README.md - Quick reference for humans

How It Works

The integration respects Aider's design philosophy:

• The AI suggests bd commands
• You run them via /run
• You maintain full control

Example workflow:

You: What issues are ready to work on?

Aider: Let me check. Run:
/run bd ready

You: /run bd ready

[Shows available issues]

You: Let's work on bd-42

Aider: To claim it, run:
/run bd update bd-42 --status in_progress

Documentation

Complete guide available at: docs/AIDER_INTEGRATION.md in the repo, covering:

• Installation and setup
• Usage patterns
• Comparison with autonomous agents
• Troubleshooting
• Example sessions

Two Philosophies, Both Valid

After implementing this, I've come to appreciate that both approaches have merit:

Aider (Human-in-the-Loop)

• ✅ User has explicit control
• ✅ No surprises from autonomous actions
• ✅ Learn commands as you go

Claude Code (Autonomous)

• ✅ Faster workflow
• ✅ Fewer interruptions
• ✅ AI handles routine operations

Beads now supports both! Choose based on your preference.

Available Now

This is live in the main branch. To try it:

# Update bd
go install github.com/steveyegge/beads/cmd/bd@latest

# Set up aider integration
cd your-project
bd setup aider

# Start using it
aider

Help & Feedback

If you run into any issues or have suggestions for improving the Aider integration, please let me know! I'm happy to help troubleshoot or enhance the integration based on real-world usage.

Thanks again for bringing this up - it led to a much better understanding of different AI workflow philosophies and a more inclusive tool design.

---

Related:

• PR: commits 89a5752 and 210789f
• Issue: bd-3djj (closed)
• Docs: AIDER_INTEGRATION.md

[kevglynn]

I run beads in Cursor daily and have onboarded a good number of engineers onto this workflow. Here's what actually works.

The thing that made the biggest difference wasn't AGENTS.md prompting. It was exposing bd operations through a local MCP server. Instead of the agent shelling out to bd ready and parsing stdout, it calls find_ready_work as a native tool and gets structured JSON back. Same for list_beads, search_memories, etc. The agent treats beads as first-class project state rather than an external CLI it's being told to use.

The other high-ROI piece: .cursor/rules/*.mdc files that are always-applied. Mine prevent the most common agent failures: opening interactive editors (bd edit), creating markdown TODOs instead of beads, forgetting --json. The rules fire on every message without eating prompt context.

I also run a knowledge persistence layer (.beads/memory/knowledge.jsonl with auto-recall at session start), which prevents the "agent re-investigates a problem it solved 3 sessions ago" loop. But the MCP integration alone is the 80/20 move.

Happy to share more specifics on the MCP setup if anyone's doing Cursor + beads.