Core banking software designed for the frontier.
LLMs are disrupting everything. When harnessed effectively, AI tooling accelerates development, increases security, and improves operational efficiency across the board. Galoy helps banks wield the power of artificial intelligence from core to customer.
The future of AI-native banking
AI is changing how software is written, how bank administrators manage day-to-day operations, and how customers interact with their accounts. The Galoy Core was designed with all three in mind, ensuring each user or actor has access to exactly the things it is supposed to have, and every action is authorized, logged, and reviewable.
Sovereignty in the age of AI
A bank built with decentralization and modern AI tooling at its core has structural advantages over incumbents needing to retrofit pre-AI architecture. And a bank that depends on a single hosted LLM ties part of its operations to that vendor’s decisions. The Galoy Core is built so hosted and self-hosted models can each be used for what they do best, and swapped as better ones arrive, without tying the bank to any single vendor.
Drua: the open-source harness for building banks
Building and managing mission-critical infrastructure with LLMs requires a control plane: something that decides what an agent may touch, keeps credentials out of its hands, and records everything it does. Drua, the banking-specific AI harness developed by Galoy, enables interoperability between self-hosted and hosted LLMs to provide resiliency and robustness for admin and customer operations. Its source is public at github.com/GaloyMoney/drua.
Credential isolation
Agents run in disposable sandboxes that never hold a real credential. Code is cloned through Galoy’s own git proxy, which authenticates the sandbox with a short-lived identity and keeps the real token on the other side of the boundary.
Human approval gates
Agents write to bot branches only. Merge is absent from the tool list entirely, so no agent flow can discover it. A human reviews and merges, and every request is recorded with the identity it acted under and on whose behalf.
No single model
Routing is provider-agnostic, with primary and fallback chains configurable per role and per workflow step. Nothing in the system depends on one lab’s availability.
Progressive disclosure
Connecting an agent to CI, observability, source control, and Kubernetes normally means hundreds of tool definitions in every request. Drua exposes three (search, describe, call) and the model discovers the rest on demand.
Composed calls
Instead of five tool calls across five turns, an agent writes one short script that runs in a restricted JavaScript sandbox and filters the results before they reach the model, so the model receives the answer it asked for.
Output that summarizes itself
An oversized tool result is elided in place, stored, and returned as a summary with a recovery handle. The agent pulls back exactly the slice it needs: a byte range, the last lines, a window into an array.
Semantic code search
Galoy’s repositories are chunked along real code boundaries (functions, structs, traits), embedded locally, and rebuilt nightly. It also mines past review comments, so an agent can search for anti-patterns the team has already flagged.
Declared workflows
A repeatable job is declared once as a YAML workflow: agent steps, deterministic steps that call no model at all, and wait steps that park a run until a webhook matches. Every step declares an output schema.
Galoy uses Drua to build the Galoy Core. It is not a commercial product. The source is public so a bank evaluating our engineering can read exactly how we govern the agents that work on it.
There is more to say about building with AI than fits on one page. If you want to hear it, get in touch.
Build with Galoy
Galoy engineers technology for leaders who are redefining money and banking.
If that’s you, get in touch