The console is at proxium.tech/app. Each screen shows the project that you select in the top bar, or all your projects. The screens that show traffic have a Time window: Today, Last 7 days or Last 30 days. Select ? beside a title to read what the screen or the panel shows.
| Screen | Use it to | Explained in |
|---|
| Overview | See the spend, and cap an app | Track spend |
| Endpoints | Get a base URL and a key for your apps | Create endpoints and keys |
| Requests | See each attempt, and why calls failed | Debug a failed call |
| Performance | Compare how fast, how often limited and how costly each vendor and model is | Monitor your project |
| Usage | Read the tokens, the cost and the calls over time | Track spend |
| Classifier | See how the classifier decides, how each request for auto got its tier, and write what the classifier asks Jev or Clef | Automatic routing |
| Memory | Review what the project learned | Use project memory |
| Memory settings | Turn memory on, set the retention and the goal check | Use project memory |
| Loops | See each conversation that counts now, and each run that finished | Limits |
| Routing | Set the models of each tier | Route requests to models |
| Keys | Make a plain virtual key | Create endpoints and keys |
| Providers | Add vendors and their keys | Add vendors and your own keys |
| MCP servers | Give your agents the tools of an MCP server | Add MCP servers |
| Data protection | Flag, redact or block sensitive data | Protect sensitive data |
| Secrets found | See the secrets that people pasted into stored conversations, and dismiss them | Sensitive data |
| Plan reuse | Let an agent's planned tool calls skip the model | Plan reuse |
| Improve | Find what to change to spend less | Lower your costs |
| Team | Add and remove people | Invite your team |
| Settings | Set what Proxium stores, export or erase the project | Data handling |
| Billing | See your plan | — |
Settings, Team and Billing open from the top bar. The other screens are in the strip, in four groups:
| Group | Screens |
|---|
| Watch | Overview, Usage, Explore, Requests, Performance, Memory, Loops |
| Configure | Endpoints, Routing, Classifier, Plan reuse, Memory settings, Providers, Keys, MCP servers |
| Protect | Data protection, Secrets found |
| Improve | Improve |
All projects
If you are a member of more than one project, the project list in the top bar starts with All projects. The console opens on it. The address and your browser keep your choice.
| Screen | In All projects |
|---|
| Overview | The totals of your projects, and Projects: the spend, calls, failed calls and last use of each project. A project that cannot route a call shows cannot route |
| Requests, Classifier › Log, Improve | The data of all your projects. Each row shows its project |
| Performance, Usage | The models and vendors of all your projects together. Usage adds Project to Split by |
| Data protection, Settings, MCP servers | The matches, the alerts sent and the tool calls of all your projects, and the list of your projects |
| Endpoints, Routing, Classifier › Diagram and Setup, Plan reuse, Secrets found, Keys, Providers, Memory settings, Team, Billing | The list of your projects. Select one to open the screen for it |
Select the name of a project in a row to open that project on the same screen. A change applies to one project only, so you make it in the project.
Overview
| Element | Shows | You can |
|---|
| Needs attention | The list at the top, worst first, at most five lines. A line is one rule that crossed its line: the budget at 80% or more spent, a vendor that answered fewer than 95% of its calls, a model that was rate limited above 5%, a model whose p95 doubled from the first half of the window to the second, or a retired model. Each line opens the screen behind it. A rule needs 20 attempts before it speaks, so one bad call raises no line. Without a line that crossed, the list says that nothing crossed a line | Open the screen of a line |
| Spent | The money of the window. Calls on a plan are left out; the line under it names what the plans used at list price | |
| Remaining | ∞ during the open beta. In All projects, Projects shows the number of your projects | |
| Saved by cache | What the cache saved, less what you charged for hits | |
| Requests, Tokens | The calls and the tokens of the window | |
| Spend over time | The cost for each time bucket | |
| Model split | Each model by cost, with its share and its tokens | |
| Vendors | Each vendor: answered calls, input tokens, cache reads, output tokens, all tokens, cost (with the calls on a plan) and its share of the cost | |
| Per-app burn | The spend, share, tokens and calls of each x-proxium-source | Set Calls per hour and Dollars per day for one app |
| Per-key burn | The spend of each key | Select a key to see its apps |
| Projects | In All projects only: each project, busiest first | Select a project to open it |
Endpoints
| Element | Shows | You can |
|---|
| Endpoints, With own routing, Expired | The counts of your endpoints | |
| Base URL | https://proxium.tech/v1 | Copy it |
| Your endpoints | Each endpoint: name, routing, last use | Routing, Revoke |
| New endpoint | | Create an endpoint, with its own routing |
Requests
| Element | Shows | You can |
|---|
| The app bar | The calls of each app | Show one app only |
| Requests | One row for each request: When (with a short id under it), Route (the classifier hops and each model tried, with the reason of a failure in words), Cost and Result (answered, failed over, failed). Select a row for the app, the kind, Tokens (input · read from the cache · output), Took, the whole request id, the road of the request, and the prompt and the answer. A cost with + is a lower bound: a model of the request has no price | Show All calls, Your apps, Proxium memory, Failed or failed over or Lost the cache, and sort them by Newest or Slowest. Select a row for the road of the request and the prompt and the answer |
The same traffic seen from the side of the vendors: which model is fast, which one limits the rate, and what a call costs. The window is Today, Last 7 days or Last 30 days. The screen opens with one sentence: the fastest model, the cheapest call and the most reliable model, of those with at least 20 streamed calls.
| Element | Shows | You can |
|---|
| Latency band | The busiest models over time, as a band from p50 to p95 | Show Answer time or First token, and pick the models |
| Rate limits over time | The share of the attempts of each model that were answered 429, for each time bucket | |
| Speed against cost | Each model as a bubble: p95 against the cost of one answered call, sized by its calls | |
| Failures | How the attempts of each model ended over time, and the causes of the failed attempts in words, biggest first. One cause across several vendors is one line | Draw Attempts or Failures over time |
| Failures by provider, Failures by reason | The failed attempts of each provider with its share of them, and the failed attempts by provider, outcome and reason with their models | |
| Models and vendors | By vendor and By model: answered calls, the share of the calls sent that were answered, input tokens, cache reads, output tokens and cost, as each vendor billed them, and the failures. A vendor row adds the mean time; the calls on a plan sit under the cost. A model row adds p50 and p90 and its state. Allowances: each vendor request or token allowance with what it used and when it resets. Retired now: each retired model and vendor, why, and until when | Open a row for Retire, Retire vendor, Bring back and Bring back vendor |
| Answer health | How the attempts of each model ended: answered, empty, cut at max_tokens, or failed. It counts what the gateway saw and does not say whether an answer was right | Show it as bars or as a series over time |
| Failover flow | The name the caller used and the model that answered, with the requests and the share of each pair. A model name on the left that differs from the answering model is a failover | |
| Vendor scorecard | One row for each of the ten models with the most calls, then every model of the window: Answered (the share of the attempts sent that the vendor answered), Time to answer p50, p95 and p99, First token p50 and p95, Speed in output tokens per second, 429 (the rate-limited attempts and their share), Calls with the output tokens, and Cost of one call with the total under it. The best value of each column is marked. The heading row holds at the top of the panel and the model column at its left while the body scrolls | Sort by p95, First token p95, Speed, Answered or Cost per call, and press Show all 30 models to draw every model of the window (the same control returns to Show the 10 busiest) |
Usage
The tokens, the cost and the calls of the project over time, split by what you choose. The MCP tool proxium_read_usage_series answers the same numbers.
| Element | Shows | You can |
|---|
| Time range | One hour, 24 hours, 7, 30 or 90 days. A range up to 3 hours is drawn by the minute, up to 8 days by the hour, up to 120 days by the day, and a longer one by the week | Pick the range |
| Split by | Model, Vendor, App, Key or Tier, and Project in All projects | Pick the split |
| Tokens, Cost, Calls | The totals of the range, with the Busiest minute and the share of the input that the vendors read from cache | |
| The chart | The metric you chose over time, with one series for each group | Draw Tokens, Cost or Calls |
| Tokens by kind | New input, cache reads and output, for each time bucket | |
| By model, By vendor, … | The twelve groups with the most tokens, and the rest as (other): tokens, new input, cache reads, output, cost and calls | |
| Cache | The hits and what they saved, for each time bucket, by layer and by model, and the last hits with the caller that sent them | |
| Who calls | The attempts and failed attempts of each app. Your apps are the tools and services that call with your keys or your login. Proxium, for this project is the memory worker, which finds facts in your conversations and searches them for each call | |
Classifier
One screen with three tabs. The tab is in the address (#/classifier?section=diagram, log or setup), so a link opens the tab that it names. The old addresses #/classification and #/routing?section=classifier open the same tabs.
| Tab | Shows | You can |
|---|
| Diagram | Five boxes from left to right: Reuse (the share of the requests that reused the decision of their conversation), Facts (what the document reads from the request), Questions (each question, its kind, what it reads, and whether it is asked once for each item of a list), Rules (each rule in words, what it sets, and in how many of the classified decisions it held; the counts start on 16 Nov 2026), and Decision (what a call gets, and the tiers it can pick) | Select the time window, and open Change the document |
| Log | How each request for auto got its tier. The figures: Requests for auto, Reused the decision (the share that read the decision of its conversation), Classified now (System One or a language model decided it), Timed out → standard (no hop answered in time), Classifier calls failed (of all classifier calls), and the typical time of the hop that answers most. Each hop reads as vendor/model | Select the time window, and read the decision of one request |
| Setup | What the classifier asks Jev or Clef, and what the answers do | Write the document of a scope, and test it |
Log
| Element | Shows | You can |
|---|
| Requests, Decided by Jev, Response time, Failed calls, Fallbacks | The figures of the window | Select the time window |
| Routing | The requests for auto, who decided them, and the share of each tier | |
| Response time | The typical time, the time 95% of decisions stay under, and the slowest answer of each classifier model | |
| Confidence | How sure Jev was, and the average confidence of each tier | |
| Models | Each classifier model, its calls and how many it answered, and each call that failed | |
| Top projects | In All projects only: the projects that call the classifier most | Select a project to open it |
| Decision log | The newest decisions: app, decider, confidence, tier and time | Select a decision to read its route, or Details for the request |
Setup
One page for the whole call of the classifier. The bar at the top stays in view while you scroll. The page has four parts, in the order of the work.
| Element | Shows | You can |
|---|
| Applies to (bar) | The project, or one app or key, with the version of its document | Select a scope, or add an app (its x-proxium-source) or a key (its prefix). An app wins over a key, and a key wins over the project |
| Classifier (bar) | Jev or Clef. The page opens on the first model that your project's classifier chain asks | Switch the model. The page then offers only what that model takes: Clef has no box question, no long context, and needs two or more options or levels |
| Unsaved changes, Save, Versions, Remove this document (bar) | The state of the draft | Save, open an earlier version as a draft, or remove the document of the scope. Save stays off while the page lists a problem |
| What should the classifier decide? (a scope with no document) | Checks to tick, in three groups: Routing, Context and Safety. Beside them, the document that applies now, in words | Tick the checks and create one document. All questions go to the classifier in one call. Or copy the document that applies, or write your own |
| 1. What the model reads | The parts of the request that go into the call | Switch on any of these parts: the last message, the messages before it, the system message, the whole conversation, the output limit, the app, the project memory, a header |
| 2. Questions | The questions that the model answers in one call | Add a question: Yes / no, Pick one, Rate on a scale or Find in an image. Add a ready-made check. Open a question with Edit to write what the model judges and to list its options or levels |
| 3. Rules | One sentence for each rule: "If … , …" | Open a rule with Edit. Add tests that fit the type of the question. Add actions: use a tier, try a model first, set the reasoning level, add the project memory, keep the cached model, flag the request or block it |
| Advanced | When the next classifier answers, a chat model as the last step, the call options of the model and the document as JSON | Edit any of them |
| 4. Test (right) | The decision, which rule did what, what the model answered, and the call that was sent | Run the draft on a request. A test is one paid call to the classifier. A test of an earlier draft shows that you changed the document |
A scope with no document uses the document of the next scope, and at the end the template Tier and reasoning. Start from a template, from a copy of the document in force, or from nothing.
Memory
| Element | Shows | You can |
|---|
| Knowledge page | What the project knows, written from its memories | |
| Memories | The memories, to review first | Confirm, Pin, Edit, Remove, add a New memory |
| Recent conversations | The kept conversations, with title and summary | Search |
With memory off, the screen says so and links to Memory in Configure.
Memory settings
A setting of one project, so it is in the Configure run of the strip.
| Element | Shows | You can |
|---|
| Setup (top) | Memory on or off, the memory cost of this month, and the goal check with its share, daily cap, credit and judge | Turn memory on, set Only these senders, Delete conversations after and the Goal check |
| Erase one user (bottom) | | Erase one end user. Owners only |
Loops
A loop is an agent that turns more than once. The loop is on for a project that set nothing. The gateway counts every conversation and refuses no call, because every ceiling of that project is 0. A project stops counting with Stop counting on this screen or in Settings › Loop, and a call turns it off with the header x-proxium-loop: off.
A conversation is one task. The gateway takes the session anchor of the call. A call that carries the x-proxium-conversation header joins the conversation that the header names. The header names a conversation, or groups the calls of several sessions. It does not decide whether the project counts.
| Element | Shows | You can |
|---|
| Agent runs | Running now, Cost so far, Finished and Goal reached, and who approves a new goal of a running agent | Stop counting or Start counting, choose who approves a new goal, open Ceilings |
| Goals that wait for you | Each new goal that waits for a person | Approve, Reject |
| Running now | Each conversation that made a call in the last ten minutes: its short id (a link to Memory), app, calls, cost, repeats and last call | Open the conversation |
| Finished | Each run that ended: calls, cost, how it ended, goal and end time | Open the conversation |
- Calls is the number of calls of the conversation, and Cost is what they cost in total.
- Repeats is the number of repeated tool calls. It counts a call with the same arguments, or the same error, when the work did not change between the two.
- When a conversation is over its ceiling, the gateway refuses the next call with
429 loop_capped. Raise the ceiling in Settings › Loop, or set Over a ceiling to alert to let the calls pass.
- The screen shows one project. Over All projects it asks you to select a project first, because a conversation belongs to one project.
- See The loop for the settings and the routes.
Routing
| Element | Shows | You can |
|---|
| Which routing to show | The default chains, or the chains of the project | Switch the view |
| Ready-made routing | The preset of the project, its version, and the models of each tier | Use this preset, Copy and edit, Reset to the preset |
| Text routing, and one section for each other type | Each tier, and Resolves to: its chain | Change, Use default |
| Where a call to this scope goes | The drawing of the routing as it runs: the scope that sets each chain, the model that a call reaches first, and every hop after it | |
| Add a hop under a chain, and the arrow and close buttons of a hop | The chain as you edit it | Add, move or remove a hop, then Save as version N, or Undo my edits |
| Versions | Each version of the document of the scope | Go back to it |
| Scoped rules | The rules for one app or one key | Add a rule |
| When a call fails | The failover policy: extra calls to one model, first wait, what a 429 does, longest Retry-After wait, how long a rate-limited vendor and a vendor out of quota are skipped, the retries of a cut answer, the small output budget, models in a plan, the catalog fill, the rules that retire a model or a vendor, and the alert | Save, Reset to default |
| Model catalog | Each model, with Cost, per Mtok | Filter the catalog |
The drawing is also where a person edits a chain, and each edit shows where a call would go before a save. See Edit the routing in the drawing.
Keys
| Element | Shows | You can |
|---|
| Your keys | Each key: name, prefix, limits, status, last use | Limits, Revoke |
| Limits of a key | Calls per hour, Requests per minute, Tokens per minute, Dollars per day | Save limits. An empty field is no limit |
| Name this key | | Create a key. Your new key shows it one time |
Providers
| Element | Shows | You can |
|---|
| Vendors | Each vendor: name, URL, type, models, key | Add a vendor, Models, Remove |
| Your own keys | Your keys for the platform providers and for your own vendors | Add a key, Remove. Without pooled keys, the form lists only your own vendors |
MCP servers
| Element | Shows | You can |
|---|
| Servers | Each MCP server of the project: name, URL, tools | Add a server, set its credential, turn tools on or off. Owners only |
| Connect an agent | The MCP URL of the project | Copy it |
| Tool calls | Each call: when, tool, caller, outcome, time | Older calls |
Improve
The page opens with one sentence. It gives the dollars a month that the findings with a saving add up to. It also gives the count of the other findings, which fix a failure or show where the cost sits.
| Element | Shows |
|---|
| Findings, Worth doing first, Worth a look, Worth fixing | The counts, and the estimated saving a month from the findings that name one |
| Reliability, Empty answers, Wrong model names, Classifier chain, Cost, Cache, Other findings | Each group of findings with its count, what the group costs you, and the screen where you fix it |
| Paused vendors and models | One folded table of the vendors and models that the failover policy paused: the vendor, the model, why, how many times in 7 days, and until when |
Each finding shows its severity, its sentence, and one sentence of the evidence that it comes from. A vendor keeps its short name, without the prefix of the project. Classifier chain holds the hops of the classifier that answer almost nothing. Other findings holds a kind of finding that this page has no group for yet. When a vendor refuses the model name, the title of the failure names the cause. The cause gives the model that a chain asks and that the vendor does not list, and the screen to fix it.
Team
| Element | Shows | You can |
|---|
| Add a colleague | | Add a person as Member or Owner. Owners only |
| People | Each person, role and status | Remove. Owners only |
Data protection
| Element | Shows | You can |
|---|
| What it looks for | Each class: what it finds, its action, its alert, the action in force | Set off, flag, redact or block, and Alert. Save |
| Your terms | Your words and phrases, one a line | Edit, then Save |
| Try a text | Each match of a pasted text, and the text as the vendor would get it | Check |
| Matches | Each match: when, class, action, where, times, app, key. Never the value | Select a time window |
Secrets found
Memory scans each message that it stores for keys, tokens and passwords (proxium_security_summary answers the same findings). A finding never holds the secret: only its kind, its place and a preview with the middle cut out.
Change a secret at its provider first (rotate). Then dismiss the finding here. A dismissal only hides the finding, and the text of the conversation still holds the secret.
The page counts the findings that need a person first. The scan closes a value that is a URL path or a documented example at once, with its reason.
| Element | Shows | You can |
|---|
| Likely real | The open findings of the window, then Examples and URLs (closed, with the reason) and Kinds of secret found | Select 7, 30 or 90 days |
| A kind of secret | The kind, its open findings, three examples, and a button that closes every open finding of the kind as not a secret | Press the button, then confirm |
| By conversation | Each conversation with a finding: its first line, its app, its kinds, its start and its open findings | Open the conversation in Memory; Show and dismiss |
| A finding | The message, the kind, the preview, the time, and the state with the reason of a closed finding | Rotated or Not a secret |
A dismissal covers every finding of the same secret.
Plan reuse
| Element | Shows | You can |
|---|
| What plan reuse does | What plan reuse is, and the mode of the apps with no setting of their own | Choose Off, Measure only or Use |
| Apps | Each app with its calls of the last 7 days, its mode, its state and why it has no plan | Choose the mode of one app, Start measuring |
Settings
| Section | You can |
|---|
| Stored prompts and answers | Set What to store: off, errors or all |
| Cache | One line: the cache on or off, and what it saved in the last 30 days. The lifetime opens Keep an answer for, and Match a question in other words where the project has it |
| Platform access (operators only) | Allow or stop the use of the pooled keys, the platform classifier, and the platform Jev of the cache, the security checks, moderation, rerank, classify and the goal check of memory. The MCP tools proxium_read_platform_access and proxium_set_platform_access do the same |
| Spend alerts | Add Spend levels, set Warn at for the limits of keys, set the Webhook, Slack, Discord and Telegram channels, Send a test alert |
| Alerts that went out | Read each alert: when, what about, the channel, sent, failed or blocked |
| Send call records | Add a webhook or an OTLP/HTTP destination, Test, Pause, Remove. Each row shows the records sent, failed and dropped |
| Export everything | Download the data of the project |
| Erase this project | Delete the project and all its data. Owners only |
Billing
Plan shows open beta · free. Proxium is free during the open beta.