As the influence of coding grows as the biggest killer app in the generative AI space, competition is intensifying among major companies seeking to expand their share in coding AI. Against that backdrop, some companies that use AI heavily are drawing attention for building and using their own internal coding AI tools.

The Information recently reported that Coinbase, Shopify and Ramp have developed their own coding AI agents and are offering them to employees as alternatives to expensive coding AI tools provided by OpenAI or Anthropic.

Coinbase, the largest U.S. cryptocurrency exchange, developed its AI coding agent Forge in April and distributed it to all engineers at the company.

Engineers can access Forge through Slack, GitHub and an in-house web interface. Forge can help brainstorm ideas for new software features or products and also carry out the work of developing them. It can also work within Coinbase's code repositories as well as enterprise software systems such as Datadog and Sentry.

Coinbase developed Forge directly to reduce AI costs. Forge uses a model router developed by Coinbase that enables it to draw on various models as needed, including Anthropic, Google and open-source models.

Use of Forge is rising rapidly internally and is leading to cost savings for Coinbase, The Information reported, citing Chintan Turakhia (친탄 투르키아), a senior director of engineering at Coinbase. Turakhia said, “At one point we effectively had no limits on AI costs, and spending increased. Now token usage keeps rising, but spending has actually fallen.”

Coinbase CEO Brian Armstrong in June also shared on social media platform X (Twitter) how it cut AI costs by nearly half while continuing to increase token usage.

ㆍ[Tech Insight] Cutting costs in half while increasing AI use... AI token engineering is coming

The key is not usage limits but better base model settings, routing and caching. First, Coinbase is running experiments to set open-weight models such as GLM 5.2 and Kimi 2.7 as defaults in its LLM gateway instead of lowering usage caps.

Coinbase also built a system that first analyses prompts and, based on cache status and model pricing, automatically assigns them to the most suitable model. Armstrong said, “High-performance models are needed for tasks that establish complex plans, but cheaper models are enough for simple execution. Ultimately, the goal is to automate the task of selecting models itself using AI.”

Coinbase is also actively using a method that makes all requests cache-aware so it can reuse existing caches as much as possible. A cache is a temporary repository that stores previously processed prompts and answers and, when the same or similar request comes in, retrieves the stored answer immediately without recalculating.

Developing Forge does not mean Coinbase has stopped using external AI coding tools such as Anthropic's flagship AI tool Claude Code. Claude Code is the most used AI coding tool among Coinbase's 2,500 engineers.

Shopify, a large e-commerce solution platform company, also developed its own coding AI agent River in addition to Claude Code and OpenAI Codex. According to company executives, 75 percent of Shopify employees were using River as of May. River is highly efficient in token usage and can perform more coding work at lower cost, The Information reported, citing company executives.

Ramp, a fintech startup for corporate cards and spending and finance management, also developed its own coding AI tool Inspect.

According to The News Stack, Inspect integrates with Datadog, Sentry, LaunchDarkly and Buildkite and can be used without being tied to a specific model. It supports various state-of-the-art models alongside MCP servers, custom tools and proprietary workflows.

In-house AI development does not always lead to cost savings. The opposite can also happen. Bloomberg reported that Walmart limited internal use in June after usage of its in-house coding agent Code Puppy surged.

According to The News Stack, companies' in-house coding AI agents integrate with internal systems and handle asynchronous engineering workflows that verify changes based on production telemetry data.

Commercial assistants still dominate interactive development sessions where developers work directly inside an editor or terminal. The News Stack reported that the key is not whether to develop an AI coding assistant in-house or buy one externally, but whether to own the orchestration layer that sits between developers and foundation models.