For developers running coding agents

See every request your coding agents make, then choose where each one runs. No new login.

Agents 0 of 4 on
  • Claude Code Capture verified
  • Codex CLI Capture verified
  • Cursor Inconclusive
  • Claude Desktop Not tested

All Off. rflectr is not in any agent's path.

Last record
agentprovidermodellatencytokenscost basisusage window
No requests yet

The control surface as designed. Every agent starts Off. Try it.

Off Every agent starts untouched.
127.0.0.1 The only address it listens on.
Metadata only Model, latency, tokens, cost. Never the prompt.

rflectr records every request your coding agents make and lets you route any one of them to a different provider. Your login stays yours.

Runs on your machine. Prompts and responses never touch disk.

How it works

Install. Turn one on. Read the record.

Your terminal
$ claude "fix the failing test in auth.ts"
reading src/auth.ts
sending request
editing src/auth.ts
rflectr: not in the path, no record

Runs Claude Code. Its lines never change. The last line is rflectr's.

rflectr 0 of 4 on
  • Claude Code Capture verified
  • Codex CLI Capture verified
  • Cursor Inconclusive
  • Claude Desktop Not tested

Claude Code: Untouched. Not in this agent's path.

Last record No requests yet
  1. Step 1

    Install it

    Windows 11 now, macOS 14 soon. Every agent starts Off.

  2. Step 2

    Turn one on

    Pick an agent. Observe watches it. Route sends it where you say.

  3. Step 3

    Read the record

    One row per request. Model, latency, tokens, cost.

Your agent keeps its own login, interface and tools throughout.

Join the waitlist

The mode switch

Three modes. One rule.

Set per agent. The agent never knows.

Mode for Claude Code
Your agentrflectr127.0.0.1Provider Your agentrflectr127.0.0.1Provider
Untouched. Not in this agent's path.

Click a mode, or watch it cycle.

The rule

Other tools ask you to log in again, or swap the agent's session for theirs.

rflectr never edits, copies or replaces a credential or session file. That is why the switch is silent.

Where this is going

A receipt. Then a gate. Then a kill switch.

Phase one is Observe: what your agent did and what it cost, one record at a time.

The record is the asset. Once it exists, the next two steps are small.

  1. Phase 1 Windows beta

    Observe

    A record per request. Cost, tokens, model, latency, usage window.

  2. Phase 2 Planned

    Gate

    Stop a request before it leaves. A spend cap. A loop. A stall.

  3. Phase 3 Planned

    Sever

    Cut a run that crossed the line, with the record showing why.

Gate and Sever are not built yet.

Join the waitlist

Safety

It never touches your login.

Design rules, not settings. Each one names the mechanism.

  • 01

    No credential edits

    It never edits, copies, rotates or replaces a credential or session file.

  • 02

    Local only

    It listens on 127.0.0.1 and nothing else.

  • 03

    No system changes

    No root certificate, hosts file edit or system proxy. Anything like that is a separate opt in.

  • 04

    Nothing stored

    Prompts, responses and tool payloads are never written to disk. Only metadata.

  • 05

    Keys stay bound

    A key you add goes only to the provider you bound it to.

  • 06

    Failover is opt in

    Off by default. One hop, only on an exhausted quota, only before the first byte.

  • 07

    No bypass

    It never gets around billing, rate limits or safety rules. Route runs on your keys, under the provider's terms.

One machine, one person. Not a shared proxy for a team or CI.

Join the waitlist

Compatibility

What is tested. What is not.

Evidence as of 4 September 2026, from the Windows build. No claims beyond it.

AgentCaptureRoutingNotes
Claude CodeVerifiedNot yetCapture proven. Routing adapter not shipped.
Codex CLIVerifiedNot yetCapture proven. Text-only route path exists, not shipped.
CursorInconclusiveNot yetTraffic seen, capture not confirmed. Do not count on it.
Claude DesktopNot testedNot yetOpen.
  • Windows 11 Debug beta. Unsigned, for named evaluators.
  • macOS 14 and later In progress.

Windows beta

Join the rflectr waitlist.

Leave your email and platform. One email when the beta opens. Nothing to pay or install today.

Platform
Agents you run

One email when the beta opens. Nothing else.

FAQ

Before you install.

Do I have to log in again, or give rflectr my account?

No. Your agent keeps its own login. rflectr never reads, edits or copies a credential or session file.

Does it need my API keys?

Only for Route, and only for the provider you bind the key to. Off and Observe need no keys.

Does it install a root certificate or change my proxy settings?

No. Not by default, never silently. Anything like that is a separate opt in.

Does it work with Cursor?

Not proven. Traffic was seen in testing but capture was not confirmed. Claude Code and Codex CLI have verified capture.

Is it Windows only?

For now. The Windows 11 beta is unsigned and goes to named evaluators. macOS 14 is in progress.

Can it get around a provider's rate limits or billing?

No. Route runs on your own keys under that provider's rules. Failover, if you turn it on, moves one hop and only after a quota is exhausted.

What does it cost?

Not set yet. The waitlist hears first.

Is it open source?

Not decided.