Release v0.15
Build AI Applications Without Writing Code
Today we’re shipping v0.15, and it’s the release where Smarter stops being a platform you write code against and becomes a platform you just describe. An AI application that searches the web, queries your database, calls an MCP server, and defends itself against prompt injection can now be written in one short YAML file, and deployed with one command: smarter apply -f my-ai-application.yaml.
Here’s what’s new.
Guardrails: safety controls, not suggestions
System prompts are requests. A model can ignore them. v0.15 introduces manifest-declared Guardrails — deterministic software controls that inspect every message on its way to the model and every reply on its way back. They moderate unsafe content, redact personal data and leaked secrets, and stop prompt injection and jailbreaks, no matter what the model would otherwise have done. Like everything else in Smarter, you add a guardrail to an application by naming it in a manifest. No code, no exceptions.
LLMClient: the whole application in one resource
The new LLMClient resource is where it all comes together. One manifest combines a model from any provider with the plugins, MCP clients, and guardrails an application needs — just by listing them by name. Smarter handles the tool calling, authentication, guardrail enforcement, and audit trail in between. This release ships with several built-in LLMClients, so you can see the whole thing working before you write your own.
New ways to reach your data
Three new plugin types landed, each one a single manifest entry:
- WebsearchPlugin — search the web and read the pages it finds.
- SkillPlugin — package instructions and reference material, as Agent Skills, written verbatim or pulled straight from GitHub, that teach a model how to do a specific job well.
- MCPClient — an experimental client that gives your application the tools of any remote Model Context Protocol server. Smarter still won’t launch an MCP server as a local process on its own infrastructure — remote only, by design.
A workbench for testing before you deploy
WorkBench is a new prompt engineering surface in the web console for trying out an LLMClient — plugins, guardrails, and all — before it goes anywhere near production traffic.
Early scaffolding for what’s next
A few resources landed in skeleton form this release, with the full implementation to follow: Orchestrator (for coordinating multiple LLMClients in sequential, parallel, and supervisor/worker workflows), LLMHost (for self-hosted models), and Vectorsearch (for semantic search over your own documents). They’re in the docs and in the codebase now, and they’ll fill out over the next few releases.
Also in this release
A handful of fixes tightened things up along the way: a SQL injection vulnerability in the plugin app is closed, stale plugin caches and manifest round-trips are fixed, and the plugin selector’s search is now whole-word and typo-tolerant. New /api/v1/cli/resources/ and /api/v1/resources endpoints round out the API surface, and the local dev environment swapped MySQL for MariaDB.