What did our LLM requests cost by model this week, and which ones failed?
Memtro reads the LLM requests your applications log through Helicone, with model, tokens, cost, latency, status, ratings and scores. It totals cost by model, lists sessions and reports usage for end users you name. It never returns the prompt or the completion. When asked and once an administrator has granted the action by name, it can rate and score requests and sessions, set properties and tags, build datasets from logged requests, and create and version managed prompts. Pointing an environment such as production at a prompt version changes live traffic at once. Deleting a prompt or dataset is permanent.
Ask in plain English from chat, your editor, a scheduled job or your own agent. The model picks the right Helicone tool, reads what it needs, and answers with the source cited.
These tools change something in Helicone. The model is told to use them only when you ask for exactly that, and an administrator grants each one by name to the teams that may use it; a personal connection acts only after its owner allows it.
Turn events on in the connection's Events tab on the Connections page, then add the address it shows as a webhook in Helicone, choosing the same kinds; the tab shows Helicone's own steps. Each event can start a job, with the event as its input. Helicone signs every delivery, and one whose signature does not check out is refused before it is read.
Events and webhooksOnce Helicone is connected, these are the kinds of things people ask. Plain questions, no syntax. If the answer lives partly in another system you have connected, Memtro looks there too.
Helicone + MemtroWhat did we spend on each model in the last 7 days?
Which Helicone requests failed in the last 24 hours?
How much did priya@acme.com cost us in the last 30 days?
Which sessions cost the most this week?
Point production for the support-reply prompt at version 4.
Dan Fletcher, engineering lead at Acme Ltd, asked what the LLM bill was by model this week. Memtro totalled cost, requests, tokens and errors per model from Helicone's logged requests. It also said whether the requests it read covered the whole week.
Maya Okafor, an ML engineer at Northwind, asked Memtro to save her revised wording as a new version of the support-reply prompt and move production to it. Memtro saved the version, then pointed the production environment at it. Live requests for that prompt used the new version at once.
The teams that lean on Helicone most, a question each of them asks, and what they connect beside it.
A record in Helicone is rarely the whole story. Connect the systems that hold the rest and Memtro checks them together, then remembers what it learned so the next question starts further along.
Using something else for the same job, or both? Each connector has its own tools; connect as many as you like.
UptimeRobot
Honeycomb
Cronitor
FireHydrant
StatusCake
GlitchTip
GLPI
Langfuse
LibreNMS
NetBox
Snipe-IT
Woodpecker CI
Zabbix
LangSmith
GrowthBook
Drata
Intruder
GitGuardian
Secureframe
ilert
JumpCloud
ConfigCat
Split
Aikido Security
Astronomer
Comet Opik
CrowdStrike Falcon
Detectify
Domotz
LambdaTest
LangWatch
LlamaCloud
Logz.io
Lunary
NinjaOne
Northflank
Portkey
PromptLayer
PRTG Network Monitor
Raygun
Retell AI
Runpod
Shodan
Spike.sh
Squadcast
Tenable
UpGuard
Uptime.com
Vanta
Zenduty
Cortex
WorkOS
Kandji
Automox
Huntress
SimpleMDM
Healthchecks.io
Rapid7 InsightVM
ZenhubIn Helicone, open Settings → API Keys and generate a new key for Memtro. Every Helicone key can also write, and Memtro acts only when asked and once an administrator has granted the action.
In Memtro, open Connections → Helicone and paste the key into API key.
If you sign in at eu.helicone.ai, set Region to eu. Otherwise leave it at us.
Ask: "What did our LLM requests cost by model this week?"
Chat in the dashboard, the OpenAI-compatible API, scheduled jobs, and any MCP client: Claude Code, OpenCode, Cursor, Claude Desktop. The Helicone tools are the same everywhere.
Only when you ask, and only what it has been granted: rate a Helicone request thumbs up or down, when it has no rating yet; add named scores to a Helicone request, when it has no score of that name yet; set a custom property on one Helicone request, replacing any value of that key; rate a Helicone session thumbs up or down; set a Helicone session's tag, replacing the one it has; create a Helicone managed prompt with its first version; save a new version of a Helicone managed prompt (not deployed to any environment); point a Helicone environment such as production at a prompt version, which live traffic then uses; rename a Helicone managed prompt; replace a Helicone managed prompt's tags; delete a Helicone managed prompt and its versions permanently; create a Helicone dataset from named logged requests; add or remove named requests in a Helicone dataset; delete a Helicone dataset permanently. Every other tool reads, searches or lists. The Connections page labels the tools that act, an administrator grants each one by name to the teams that may use it, and a personal connection acts only after its owner allows it.
A personal connection serves only your own requests. A shared connection serves the organisation: everyone, or only the teams an admin chooses. Admins can always add shared connections; other members need the permission.
Credentials are encrypted at rest. Tool results are cached briefly so agent loops do not repeat identical calls, and the facts the model extracts are saved to your organisation's memory with provenance (system, record and link) so they can be checked and superseded.
API key. In Helicone, open Settings → API Keys and generate a new key for Memtro. Every Helicone key can also write, and Memtro acts only when asked and once an administrator has granted the action.
Yes. Turn events on in the connection's Events tab on the Connections page, then add the address it shows as a webhook in Helicone, choosing the same kinds; the tab shows Helicone's own steps. A job set to start "When an event arrives" then runs once for each matching event, with the event as its input. The run treats the event as outside data, never as instructions, and anything the run remembers is kept as unverified.
Seven-day free trial, then one monthly price per organisation. Connectors, jobs and memory are never metered.
Sign in, add a provider key, create an API key and point your tools at Memtro. It takes about five minutes.
7-day free trial · card on file, nothing charged until it ends · cancel any time
