How to connect Reality Router to Cursor

30 seconds in Settings. Every Chat, Composer, and Agent call routes through RR — smart per-call model choice, full cost dashboard, no more staring at your fast-requests-remaining counter.

If you're on Cursor, you already know the drill: the $20/mo Pro plan is fine until you actually use it, then you're rationing fast requests by mid-week. Cursor's own Settings has a way out — it's called Override OpenAI Base URL — and pointing it at Reality Router turned a $47.20 bill into $5.40 on a real 528-request stretch (89% off, dashboard shown below), without changing how anyone actually uses Cursor. Four things happen at once:

• Per-call model selection — RealitySignal calibrated probabilities pick the right model for each step, so cheap models handle boilerplate and expensive ones only fire when they'll actually help.
• No fast-request meter — pay per token instead of per bucket. Kick off a long Composer session without watching a counter.
• Live cost visibility — every request, every dollar, every model choice lands in a dashboard you own.
• Same Cursor — same Chat panel, same Cmd+K, same Composer, same keybindings. Just cheaper and metered.

Here's the whole setup.

Before you start: what routes, what doesn't

Cursor's Override OpenAI Base URL is a mode switch, not a mixer. When you enable it:

• ✅ Chat, Composer, and Agent route through RR
• ✅ Cmd+K inline edits route through RR
• ⏸ Tab autocomplete stays on Cursor's proprietary model (their infrastructure, included in Free)
• ⏸ Cursor's bundled model catalog (Claude, GPT-4o etc. via Cursor's plan) is disabled while override is on

That's the right split. Tab autocomplete is Cursor's strongest feature and it's included in the Free tier anyway. The expensive stuff — long-running agent sessions, big Composer refactors, deep Chat threads — is exactly what RR is best at routing.

1. Add RR as a custom model (30 seconds)

Open Cursor Settings → Models. Scroll to the API keys section.

• Click Add Model
• Under OpenAI API Key, enable the checkbox labeled Override OpenAI Base URL
• Base URL: http://localhost:8000/v1 (or http://<your-host>:8000/v1 if RR is on a different box)
• API Key: rr-local (RR accepts anything for local instances — the header just needs to exist)
• In the model ID field, add the ones you want visible in Cursor's picker. Good starting set:
  • auto — RR chooses per call (recommended default)
  • claude-opus-5
  • gpt-5
  • claude-haiku-4-5
  • deepseek-v4-flash
• Click Verify. Cursor pings the endpoint. Green check = you're live.

Then disable any of Cursor's pre-enabled models that overlap (the docs advise this to avoid routing conflicts). RR is now your provider.

Heads up on a known Cursor quirk: the community has reported intermittent issues when Override is enabled alongside other providers. If Chat says "network error" once, hit Verify again — 9 times out of 10 it clears immediately.

2. Pick your model (10 seconds)

In any Chat, Composer, or Agent surface, click the model dropdown at the bottom. You'll see the model IDs you added, prefixed. Pick auto and let RR decide, or pin a specific one.

3. Verify it worked (20 seconds)

Ask Cursor's Chat anything trivial:

> summarize this file

Then open the RR dashboard at http://localhost:8000. Your Cursor call shows up in the Agent Activity table.

What you'll see

Here's the dashboard mid-session, with Cursor as the top consumer:

Enlarge

Total Volume: 528 requests. Accrued Expense: $5.40. Potential Cost (if everything had run on the top model in the pool): $47.20. Retained Value: $41.80 — 89% of the bill you'd have paid without routing, saved automatically. The top row of Agent Activity is Cursor itself.

Why bother

Cursor's pricing is a black box on purpose. You buy Pro at $20/mo, get a bucket of "fast requests," burn through it faster than you expected, then either wait for reset day or upgrade. Neither the number of requests you used nor the model each one actually hit is really visible to you. And any long-running Agent job is playing Russian roulette with the meter.

Route Cursor through RR and that changes:

• No caps to hit. Kick off a full-codebase audit, a batch refactor, a research crawl — RR bills per token, no bucket, no cutoff.
• You see every call. The dashboard shows model, cost, latency, success — for every request, in real time. If something spikes, you know within seconds.
• Cost overruns become impossible to hide. A runaway loop that would have silently drained your fast requests is now visible on the same dashboard, one alert away.

Same Cursor. Same keybindings. Same Chat, Composer, Agent. Just metered, cheaper, and safe to let run.

→ realityrouter.dev

Reality Router is open source and self-hosted. Metrics in this post are from a real 528-request run.