Install the context
Boost as a development dependency, then its MCP server and guidelines. The agent now reads this application, not an average one.
LaravelCloud
Realtime
A capability story
An app agents can work on.
A platform they can run.
A coding agent does better when it reads your application rather than guesses at it. Laravel now gives it the conventions, the test results and the tools in a form it can use, and Cloud gives it an API to deploy with. The decisions that matter still wait for a person.
See what the agent sees
The idea in one line The agent reads freely, calls what you register, and asks before anything that matters.
The agent's viewRead · call · ask
It can read
Conventions and docs · Boost guidelines, skills, search-docsReads
Schema, logs, last error · Boost's MCP serverReads
Test and analysis results · PAO, as compact JSONReads
It can call
Your application's tools · Laravel MCP, behind your guardsCalls
Deploys, environments, databases · the Cloud APICalls
It must ask
A tool marked Approvable · AI SDK, paused for a decisionA person
A decision can approve, reject with a reason, or edit the arguments
Agent connected · scoped
The agent's view, surface by surface
Four places an agent
meets your work.
Each first-party piece answers the same three questions: what the agent can read, what it can call, and what needs a person. Boost and PAO are development dependencies, Laravel MCP is a route in your own app, and Cloud offers an API you hand a token to.
Laravel Boost
Guidelines assembled from the packages in composer.json, skills loaded when needed, and a local MCP server inside the app. The agent queries the schema and the logs instead of inferring them from the file tree. Since Laracon US 2026 it also infers your team's own conventions.
- Reads
- Guidelines, skills, a docs search, database schema, log entries, the last error
- Calls
- Read-only queries through its MCP tools, over stdio on your machine
- Asks a person
- Everything it writes goes through your agent's normal edit and review
Boost sheet; Monitor 556 (convention inference, 29 July 2026); first tagged on Packagist 13 August 2025.
Laravel PAO
When an agent runs PHPUnit, Pest, Paratest, PHPStan, Rector or Artisan, PAO returns compact JSON or undecorated text. When you run the same command, your terminal is unchanged. New applications include it by default.
- Reads
- Counts, failures with file, line and message; PHPStan errors per file; Rector diffs
- Calls
- Nothing new. The agent runs the same commands a person does
- Asks a person
- Not applicable; PAO only changes the shape of output
First tagged 7 April 2026; announced 14 May 2026 (Monitor 572). Laravel 12 and above, PHP 8.3 and above.
Laravel MCP
A third entry point beside the web and the API. Tools, resources and prompts are classes, registered in routes/ai.php, behind Sanctum or Passport. In 1.0, a searchable catalogue keeps the advertised list the same size however many tools sit behind it.
- Reads
- Resources you choose to expose
- Calls
- Only the tools you register, validated and authorised like a controller
- Asks a person
- Your policies decide; conditional registration is checked again at execution
First tagged 12 August 2025; public beta announced 18 September 2025 (Monitor 645); 547 (1.0 catalogues, 11 September 2026: 1,431 bytes advertised at any size, against 5,843 for 10 tools and 58,585 for 100, in Laravel's own small test).
The Laravel Cloud API
Deployments, environments, databases, caches, object storage and scaling, managed programmatically. Laravel names three users for it: CI pipelines, infrastructure as code, and AI agents that provision resources or query status.
- Reads
- Status of environments, deployments and resources
- Calls
- Deploy, provision, scale, through a token you issue
- Asks a person
- Spending limits stop new provisioning at a threshold you set · who holds the token is yours to decide
Monitor 495 (resources expanded, 29 January 2026), 491 (announced for teams and agents, 2 March 2026), 556 (API, CLI and spending limits listed, 29 July 2026).
Sources: the Boost, PAO, MCP and Cloud sheets in this guide; Monitor ids as given on each panel.
Which kind of AI this is
Not the AI in your product.
The agent at your desk.
The cloud story is about AI inside the product: a customer's message, a model, a person who approves. This story is about agents working on the codebase and operating the platform. The same rule holds in both: the model never becomes the policy engine.
AI in the product
Your customers meet it. It runs in production, on your queues, with your keys.
Who triggers itA customer or an event
Built withAI SDK agents and tools
The gateAn Approvable tool
Same rule
Agents on the codebase
Your developers work with it. It runs beside them, in development and in CI.
Who triggers itA developer, or a pipeline
Built withBoost, PAO, MCP, the Cloud API
The gateReview, tests, a scoped token
Fair to sayAnthropic shipped deferred tool loading
Laravel's MCP post credits it, and an open MCP proposal, as independent arrivals at the same idea.
Laravel's distinctionA package, not a platform feature
Nothing new to operate: each batched call goes through the same invoker as a normal tool call.
“A search result is never a capability grant.”Laravel blog, on Laravel MCP 1.0, 11 September 2026 · Monitor 547
One real task, start to finish
The AI SDK v1.0 upgrade,
done by an agent.
Boost
Run the guided prompt
/upgrade-ai-sdk-v1 in Claude Code, Cursor, OpenCode, Gemini or VS Code. Each breaking change arrives with its likelihood of impact.
Read the tests as data
Pest runs as usual. PAO hands the agent counts and failures as JSON, so the next edit is aimed at a file and a line.
Run the backfill, once
The guide's backfill migration for the new steps column must run before v1.0 is deployed. Someone who knows the data should see it go.
Deploy from the pipeline
CI, or the agent with a token you issued, calls the Cloud API. Spending limits sit underneath whoever holds the token.
Sources: Monitor 7 (AI SDK v1.0 and the Boost upgrade prompt, 23 September 2026), 572 (PAO), 491 (Cloud API), 556 (spending limits). The flow is the author's arrangement of steps Laravel describes separately.
The evidence, with its caveats
Laravel measured it.
And said how far to trust it.
Boost Benchmarks ran six models through 17 Laravel tasks, checked by 315 Pest tests, with and without Boost. Laravel reports that every model improved or stayed roughly the same. It also says one run is not definitive and the results are directional.
haiku 4.5267 / 231
sonnet 4.6297 / 296
kimi k2.5298 / 270
opus 4.6301 / 275
gpt-5.3 codex313 / 298
gpt-5.4313 / 299
With BoostWithout
Harder tasks
+36tests for haiku 4.5, the largest gain reported- MCP servers, AI SDK agent loops, Pennant, Inertia shared data
- Simple routes and queue jobs rarely changed
Runs vary
7 → 17of 19 on one evaluation, for the same model, by running it again- A single run is not always definitive, Laravel says
- Majority-of-several runs under consideration
Some overhead
$0.05 to $0.20extra per evaluation at API prices, from the extra MCP calls and tokens- sonnet 4.6 gained one test and ran faster, 208s to 179s
- gpt-5.3 codex ran slower with Boost, 175s to 191s
Kayla Helmick
5.5 daysidea to live site, with no prior coding, using Claude Code and Cloud- Gallery, approval workflow, blog, events, email
- Her caveats: usage limits, agents that loop, developers still essential for payments and security
Model names, test counts and timings are Laravel's, from its published results; they are not a ranking this guide endorses. Kayla's figures are her own.
Let the agent read everything.
Decide what it may touch.
Three questions for a team bringing agents in: what does your agent guess today that it could read instead; which of your tools would you register for it, and behind which guard; and which actions, on the app or on Cloud, should always wait for a person.