Use the ap-... key printed on first server start, or ask an admin to mint one.
ai-proxy
All systems operational
Tokens Online
—
—
Headroom 5h / 7d
—
—
Requests
—
Tokens Processed
—
Total processed
Avg Latency
—
Average response time
Wall of Sh(f)ame
via proxydirect
Amber is direct use of your account; the neutral ring is use via the proxy
Token Pool
Tokens over time
Last 24 hours
Token Distribution
In-flight requests
Live
Recent requests
Time
User
Token
Model
Status
Input / Output
Duration
Quick Actions
Token Pool
All Anthropic accounts currently feeding this proxy.
Activity
Requests, token usage, and latency over time.
Tokens over time
Users
People who can use this proxy. Each gets their own API key.
#
Name
Role
Status
Cache (24h)
Reqs (24h)
Fresh tokens (24h)
Last active
Contributed
Actions
API Keys
Your personal key, and how to point Claude Code or Codex at this proxy.
Your API key
Endpoint
Works with
Claude Code and Codex. Requests are logged under your name.
Set up a client
Add to your shell rc file (~/.bashrc, ~/.zshrc)
Run claude as normal
Nothing to install, no Anthropic login needed on your machine.
Paste into ~/.codex/config.toml
Paste it at the very top of the file.
Run codex as normal
No ChatGPT login needed on your machine. The key lives in the config file, not your shell: Codex's background daemon makes the requests and never sees variables exported after it started.
Add a token
Contribute a subscription to the shared pool. Each provider has its own pool.
Anthropic
Claude Pro / Max · serves Claude Code
1
Connect with Anthropic
2
Paste the code
We'll open Anthropic's authorization page in a new tab. Log in with the account you want to share, and you'll be shown a code to paste back here.
Authorization tab opened. Log in, copy the displayed code (it looks like xxxxx#yyyyy), and paste the whole thing below.
Sign-in happens in a small window. The last step looks broken, but it isn't. Here's what to expect:
Sign in with the ChatGPT account you want to share, and approve Codex.
The window then shows an error page. That's expected: the address is what we need, not the page.
localhost:1455/auth/callback?code=ac_…
This site can't be reached
Copy that window's whole address (click the address bar, then Ctrl/⌘+C), close it, and paste it here.
Sign-in window opened. When it shows "This site can't be reached", copy its address (starts with localhost:1455/auth/callback) and paste it below. Finishing is automatic.
Token added
This is a separate login from your own Codex: the pool keeps its own session, so your local codex stays signed in.
Fallback pool
When every token in this pool is at its limit, requests from borrow capacity from the fallback pool instead of getting a 429. Borrowed requests run on other people's subscriptions.
Logs
Recent proxied requests, most recent first. Rows marked
retried are pool tokens that
failed and were retried on another token -- the client never
saw those statuses.
Time
User
Token
Model
Status
Input / Output
Duration
Usage patterns
Turn session evidence into better guidance.
Loading
Pattern totals · all sessions
Checks collected—
Awaiting review—
Review coverage—
Detection threshold—
Your review workspace
Choose a check to review the assessment and label what happened.
Check
Claude’s assessment
—Jev probability
What actually happened?
Full grading criteria
Labels save immediately. Judge what happened independently of Jev’s score.
Conversation evidence
Formatted from the saved tail. Raw text preserves the exact evidence.
Tool activity
For grading only
MCP available
Tools called
Calls reflect the retained tail. These facts never decide the verdict.
Audit details Grading facts & definition
Grading factsPattern definition at check timeExact request sent to Jev
Find a threshold you can trust.
Explore the trade-off between useful detections and missed opportunities across all sessions for this pattern. Check-list filters do not change these metrics.
Precision & recall
Exploring doesn’t change the pattern. Saving a threshold applies to future checks.
At this threshold
Early evidence
Precision measures how often detections are real. Recall measures how many real opportunities are detected. Review negative checks too.
All threshold measurements
Threshold
Precision
Recall
True positives
False positives
False negatives
True negatives
Define what matters.
Keep the pattern clear and choose whose sessions contribute evidence.
Check frequency
Applies to this instance
Choose how much conversation growth triggers a check. Token counts are estimated; changing the frequency takes effect without a restart.
Pattern definition
User hints off
Participants
Use a person’s switch to enable collection. If setup is needed, choose patterns and Save & enable. Edit each person’s selection under Patterns; Pattern definition changes the rule for everyone.
Loading collection configuration…
Welcome to ai-proxy 👋
This is a shared Claude pool. Pick a path to get started.
You can always do the other one later from the sidebar.
User stats
—
Requests
—
Fresh tokens
—
Observed direct use (lower bound)
—
Avg latency
—
Errors
Top models
Pool tokens used
Prompt cache by prefix
Prefix
Verdict
Model
Reqs
Read
Written
Hit
Blocks
No requests in the selected range.
Create user
API key for :
Shown once. Hand this to the user now — you can't retrieve it later. Use "Rotate key" in the users table if lost.