Tejas's blog

Morph: The Subagent Layer for Coding Agents

Website: docs.morphllm.com

Morph focuses on the parts of codegen that are better handled by specialized subagents: edits and search. We're turning code generation into a commodity so any AI coding product, from PR review to vibecoding platforms, can be built with a simple import.

Fast Apply: Speed at Scale

Our first model, Fast Apply, merges LLM-generated edits into real repos at around 10,000 tokens per second. In production workloads it is about 35 percent faster end to end than using Claude alone, and roughly four times faster than the fastest model on Cerebras.

Traditional code editing approaches often involve full file rewrites or failure-prone search/replace operations. Morph smooths this process by using abbreviated edit snippets that are intelligently merged with existing code.

Deterministic code merging has hundreds of edge cases that you can never fully squash. A model is a perfect use case for this. We exploit the fact that the output merged code is remarkably similar to the input code—a property we use to make our speculative decoding extremely fast.

Memory Bandwidth Bottleneck

Everyone complains how LLM variable length inference is inherently memory bandwidth bottlenecked.

My hot take is that it's just a skill issue. You're essentially admitting you don't know what to do with your extra compute. And you're too lazy to find out. Speculative decoding is a great way to tackle this, especially in scoped niche domains.

Building the Subagent Stack

We use this traffic to train small, task-specific models for code search and model routing that bigger models call as tools to search, debug, and ship changes. This subagent layer makes coding agents execute quickly and consistently in production.

Teams like Framer, Webflow, Block, Vercel and others already use Morph as the layer that makes coding agents work in production. In the last five months we went from zero to 1.1 million dollars in revenue and are heads down turning this subagent layer into the default stack for anyone trying to ship real coding agents.

Commoditizing Vibecoding

I've seen virtually every vibecoding platform's infra and stack. They're generally all similar, and a bunch of repeat work is being done.

We're commoditizing the subagents and building on top of them to make Coding Agents a simple primitive that people can use to build unique things—products like Zo.computer gizmo, websim, and more. The infrastructure layer shouldn't be where everyone reinvents the wheel. It should be a solved problem so builders can focus on creating novel experiences.

Why Specialized Subagents Matter

Large language models are powerful, but they're not optimized for every task. By handling edits and search with specialized subagents, we can:

  1. Maximize Speed: 10k+ tokens per second for code merging, 35% faster than Claude alone
  2. Reduce Costs: Task-specific models are dramatically cheaper than using frontier models for everything
  3. Increase Reliability: Models trained on specific tasks perform better than general-purpose models
  4. Enable Real Production Use: Fast, consistent performance makes coding agents viable for production workloads

How It Works

Morph provides a simple integration:

  1. Generate Code Snippets: Configure your LLM to output abbreviated edit snippets instead of full file rewrites using special comments like // ... existing code ...
  2. Merge with Fast Apply: Pass the edit snippet and initial code to Morph's OpenAI-compatible API endpoint
  3. Use Specialized Tools: Let your agent call Morph's code search and routing models as tools

Key Features

  • 10,000+ tokens/second processing speed with optimized models and speculative decoding
  • OpenAI-compatible API for easy integration
  • Task-specific models for code search and routing
  • Production-tested by leading companies
  • API base URL: https://api.morphllm.com/v1

Getting Started

To get started with Morph:

  1. Sign up for a Morph account and create an API key
  2. Try the playground to test the lightning-fast code merging
  3. Configure your LLM to output abbreviated edit snippets
  4. Integrate the OpenAI-compatible API into your workflow

For access to longer context models, self-hosting, or business inquiries, contact info@morphllm.com