Skip to content
Scalpels home

Scalpels is officially launching on October 31, 2026. Special early adopter pricing is available now.

Assay

Laravel MIT licensed Experimental

See every laravel/ai run, token count, tool call and failure without sending the record to somebody else's telemetry account. Capture content only when you choose, then turn the useful production cases into datasets you can keep and export.

Deploy Assay as a supporter Or run it yourself, free Read the Docs

Work in progress. Available to supporters.

Replaces
Hosted AI observability
Runs on
Your Laravel Cloud account
Needs
Compute · Database · Queue · Object storage · Scheduler

What it records

Usage is always recorded. Prompt and response content is a separate, explicit capture choice.

  • Provider and model usage

    Break tokens and operations down by application, environment, provider, model, agent and opaque subject.

  • Whole run trees

    Follow agents, steps, attempts, tool calls, timings, outcomes and failovers as one connected run.

  • Selective content capture

    Keep usage-only telemetry by default, or capture filtered prompts, responses and tool data for chosen traffic.

  • Review and labels

    Flag production runs with labels, ratings and notes while the evidence is still attached.

  • Curated dataset exports

    Collect selected runs into retained datasets and render the current data as versioned JSONL on demand.

  • Scoped agent access

    Query usage or work with captured content through separate MCP abilities, including two-phase subject erasure.

  • Useful before you store a prompt

    Usage mode records the shape, timing, model, tokens, failures and tool names of a run without storing prompt bodies, responses or tool payloads. Full capture is a deliberate second step, with filtering owned by the reporting application.

  • Erasure survives a restore

    Subject erasure writes a durable journal outside the database, replaces the subject with a keyed tombstone and can reapply later journal entries after a snapshot restore. The key is separate from APP_KEY so application-key rotation does not rewrite identity.

Assay vs hosted AI platforms

A narrow Laravel instrument, not an AI engineering suite

LangSmith, Langfuse and Braintrust cover more frameworks and more of the evaluation loop. Assay is for a Laravel team that wants the telemetry and curation slice on infrastructure it already operates.

Assay feature comparison
Feature Assay Assay is MIT licensed and Experimental. The server is ready to fork; the reporting package is not publicly released yet. LangSmith Hosted observability and evaluation from LangChain. Langfuse Open-source observability, evaluation and prompt management. Braintrust Hosted observability and evaluation with a self-hosted data plane option.
Laravel AI capture the row we rebuilt Does this well:Purpose built Partly, or with a caveat:Custom instrumentation Partly, or with a caveat:OpenTelemetry Partly, or with a caveat:General SDKs
Usage and run trees the row we rebuilt Does this well:Built in Does this well:Built in Does this well:Built in Does this well:Built in
Production traces into datasets the row we rebuilt Does this well:Curate and export Does this well:Built in Does this well:Built in Does this well:Built in
Automated evaluations we did not build this Not built:Not built Does this well:Online and offline Does this well:Online and offline Does this well:Built in
Prompt management and playground we did not build this Not built:Not built Does this well:Built in Does this well:Built in Does this well:Built in
Quality alerts and release gates we did not build this Not built:Not built Does this well:Built in Partly, or with a caveat:CI experiments Does this well:Built in
Framework coverage Not built:Laravel 13 only Does this well:Broad SDK support Does this well:SDKs and OpenTelemetry Does this well:Broad SDK support
Where your telemetry lives Does this well:Your Laravel Cloud Works, but on their terms:Cloud or enterprise Partly, or with a caveat:Cloud or self-hosted Partly, or with a caveat:Cloud or data plane
Cost of more telemetry Does this well:Your infrastructure Works, but on their terms:Plan and usage Works, but on their terms:Plan or infrastructure Works, but on their terms:Plan or infrastructure
Who is on call for the tool itself Partly, or with a caveat:You are Does this well:They are Partly, or with a caveat:Depends how hosted Partly, or with a caveat:Depends how hosted
  • Does this well
  • Partly, or with a caveat
  • Not built
  • Works, but on their terms

The honest test

Look at how your team investigates AI behaviour today. Moving loses little when all three of these are true.

  • The applications reporting telemetry are Laravel applications using laravel/ai.
  • You need trace review and curated data, not an automated evaluation platform.
  • Your team is willing to operate the queue, database, scheduler and bucket it owns.

Then Assay keeps the useful slice close to the code and data it describes, with no separate observability account between you and the rows.

Keep paying them if

These are the parts we chose not to build, and we are not going to pretend otherwise.

  • You need automated judges, experiments or regression gates.
  • Your team manages prompts in a shared playground.
  • Your applications span several languages and AI frameworks.
  • You want vendor support and somebody else on call for the telemetry service.

Small, sharp tools are a deliberate scope, not an unfinished roadmap. Assay will not grow into a general AI engineering platform.

Be honest about it

You don't need us to run Assay

The repository is public and MIT licensed. These are the steps, and you are welcome to spend the afternoon doing them by hand.

  1. 1

    Fork the Assay repository into your own GitHub organisation.

  2. 2

    Deploy the fork to Laravel Cloud with Postgres, a managed queue, object storage and the scheduler. There is nothing to set: Scalpels generates the dedicated erasure key into your own Cloud environment once, and Assay writes its erasure journal to the attached bucket on its own.

  3. 3

    The Assay client is not publicly released yet. While Assay is Experimental, connect a reporting application from a local Assay checkout, then mint its one-time installation credential in Assay.

Or become a supporter, and Scalpels deploys it for you. It forks it, provisions exactly what it needs on your Laravel Cloud account, and deploys it. The app and the infrastructure are yours either way — the only question is who spends the afternoon.

Deploy Assay as a supporter

Want some help?

After you buy, we will pay for an hour with one of these experts to help you get everything up and running.

  • Artisan Build

    USA

    We run Scalpels, as well as a suite of other tools for developers and development teams.

    Agency Ask for help
    Ask Artisan Build for help

Assay pairs with the rest of the suite

Hone covers the Laravel application around each AI run, Reel shows the user session that led there, and Capstan gives your local agents another self-hosted surface to work through.

Hone Reel Capstan