What GitHub tells us about AI in banking

Most of what is written about AI in banking comes from surveys and vendor announcements. GitHub offers a different lens: what people are actually building, forking and maintaining in public. It is an imperfect lens — banks do not put production systems on GitHub, and stars measure attention rather than adoption — but it is an early one. The open-source record usually shows where the engineering effort is heading a year or two before it appears in bank architectures.

The backdrop is scale. GitHub reported 4.3 million AI-related repositories in its 2025 Octoverse, and 1.1 million public repositories importing large-language-model SDKs, up 178% year on year²⁵. Within that, we looked specifically at financial services. On 1 October 2026 we queried the GitHub API for the most significant, actively maintained projects across five categories — agents, financial models and benchmarks, data connectors, governance, and financial crime — and for what banks themselves publish. Exhibit 1 summarises the repositories that matter most.

Agents: the centre of gravity

The biggest numbers belong to multi-agent finance projects. TradingAgents, which splits trading analysis into specialised LLM roles, has 109,364 stars²; ai-hedge-fund has 63,811⁴. Both are explicit that they are research or educational tools and not for real trading or investment²⁴. Their popularity tells us that developers find the ‘team of specialist agents’ pattern compelling; it does not tell us that it works in a regulated setting.

Underneath sit the general agent frameworks: CrewAI (59,233 stars), LangGraph (42,525), OpenAI’s Agents SDK (29,788), Google’s ADK (21,688) and Microsoft Agent Framework (13,882) are all being pushed to daily¹⁰⁸¹¹. AutoGen is a useful cautionary tale: it still carries 61,243 stars, but its README now states it is in maintenance mode, with new users directed to Microsoft Agent Framework, and its last push was on 15 April 2026⁹. A bank that chose frameworks by star count would have picked a dead end.

The most significant new entrant is the model vendors’ own reference implementations. Anthropic’s financial-services repository, created in February 2026, already has 38,347 stars¹. It ships ten named agents — including a GL Reconciler, Month-End Closer and KYC Screener — plus skills and MCP data connectors, and states that the agents do not execute transactions, post to a ledger or approve onboarding: every output is staged for human sign-off¹. Cloud providers publish financial-services samples too, such as AWS’s agentic value accelerator and Azure’s multi-agent banking assistant, but with 64 and 166 stars respectively²⁰ — a reminder that the material most relevant to banks is often the least starred.

Financial models and benchmarks: building outpaces testing

The financial-LLM wave of 2023 is still visible. FinGPT has 21,303 stars and is actively maintained⁵; its README shows how cheaply domain adaptation can be done, citing a LoRA fine-tune on a single RTX 3090 for about $17.25 against an estimated $4.23 million to pre-train Llama2-7B⁵. FinRobot, from the same foundation, has shifted the emphasis to agents⁶.

Evaluation is thinner. FinanceBench, one of the most cited financial QA benchmarks, releases only a 150-case open sample of its 10,231 questions and last saw a push in December 2024; its authors reported that GPT-4-Turbo with retrieval answered incorrectly or refused 81% of questions¹⁷. The PIXIU/FinBen repository has 890 stars and was last pushed in March 2025¹⁸. The open community is far better at building finance agents than at testing them — which means banks will have to build much of their own evaluation.

Connectors: MCP becomes the plumbing

The Model Context Protocol reference servers repository has 90,771 stars⁷, and finance has followed. OpenBB now describes itself as an open data platform ‘for analysts, quants and AI agents’ (73,695 stars)³; single-purpose finance MCP servers such as financial-datasets/mcp-server have gathered 2,299 stars¹⁹. Most community finance MCP servers wrap market data, trading platforms or consumer finance apps. For a bank, each one is third-party code that will hold credentials and move data — a supply-chain question as much as an integration one.

The growth curves in Exhibit 2 are steep for every category, but the shape differs. Repositories matching ‘banking agent’ went from 28 in 2024 to 1,671 in the first nine months of 2026, and ‘finance mcp’ from 8 to 1,475²⁴: the energy has moved from building models to wiring agents into data and systems. Anti-money-laundering projects grew more slowly, from 103 to 562²⁴.

Governance and financial crime: small repositories, outsized importance

The FINOS AI Governance Framework has just 109 stars¹², yet it is arguably the most bank-relevant repository in this review. Its maintainers include engineers from several large US, UK and Canadian banks¹², and version 2 catalogues 23 risks — 11 operational, 9 security and 3 regulatory — with preventative and detective mitigations mapped to the EU AI Act, NIST, OWASP and ISO 42001¹³. Governance content does not attract stars; that is a reason to look for it deliberately.

Financial crime shows the limits of open source. Real transaction data cannot be published, so the shared assets are synthetic: IBM’s AMLSim simulator (397 stars) and its Multi-GNN graph models, built for experiments on IBM’s synthetic AML transactions (123 stars)¹⁶¹⁵. Both were last pushed in September 2025¹⁶¹⁵. They are good for method development and benchmarking; they are not a substitute for a bank’s own labelled alerts.

What banks open-source themselves

Banks publish the plumbing around AI rather than models. A global investment bank’s quant toolkit has 13,034 stars²¹; other banks’ libraries for data profiling (1,591 stars), SHAP-based model validation (154) and functional machine learning (1,551) address data quality, validation and ML engineering²³. The newest signal came in July 2026, when a large US card issuer open-sourced an agentic code-security tool it had developed internally; it runs as Claude Code skills on Opus-class models and already has 1,037 stars¹⁴. Banks are now publishing agents that help them control software risk — including the risk that agents create.

Reading the signals

• Maturing: agent frameworks with vendor backing and daily commits; MCP as the connector standard; vendor reference agents designed around human sign-off¹¹⁷¹.
• Still hype-prone: autonomous trading and ‘AI hedge fund’ demos whose own authors rule out real-world use²⁴.
• Under-built: open evaluation for banking tasks, and governance tooling that turns frameworks into running controls¹⁷¹².
• Caveat: stars are not adoption. Last-push dates, maintenance notices, licence files and who maintains a project say more about fitness for a bank.