DAG Studio · Black Swan Causal Labs

What's Inside

DAG Studio is two related products.

A drag-and-drop editor and a Model Context Protocol server, sharing one identification engine. Each can be used on its own.

Surface 01

The Editor

A drag-and-drop DAG editor with an in-browser causal-inference engine. Backdoor paths, adjustment sets, identifiability checks. No install required.

Live preview Read more ↓

Surface 02

The MCP Server

The same engine exposed as a Model Context Protocol server. AI assistants can query identifiability, suggest adjustment sets, and flag overadjustment, with citations.

Source available, npm pending Read more ↓

The Editor

Drag-and-drop DAGs, with a real identification engine underneath.

Drag-and-drop DAG construction Build causal diagrams in the browser with a tactile graph editor.
In-browser causal-inference engine Backdoor path enumeration, minimal sufficient adjustment-set search, and identifiability checks, computed locally.
Validation tab with three-way concordance Live comparison against Pearl (2009) and the dagitty R package on a battery of canonical DAGs.
Linear Gaussian SEM data simulator Sanity-check inferences before committing to a study design. Assumptions documented in-app.
Light and dark themes, full mobile responsiveness Designed for cross-functional teams who reason about DAGs from a laptop or a phone.

Adjusted backdoor path highlighted on a canonical DAG depicting confounding.

Open the Editor →

The MCP Server

An identification engine an AI assistant can actually call.

Alongside the editor, DAG Studio also runs as a Model Context Protocol (MCP) server. The intent is not to have an LLM "draw a DAG for you." It is the opposite. When an analyst is reasoning about study design with an AI assistant, the AI should be able to query a real causal-inference engine to check identifiability, suggest adjustment sets, and flag overadjustment, rather than hallucinating about d-separation.

Tool surface (v1)

01 dagitty-syntax parsing

02 Identifiability analysis (backdoor paths, adjustment sets)

03 Overadjustment check (collider, descendant, descendant-of-collider flags)

04 Code generation for R (dagitty DSL) and Python (networkx)

05 Effect-modification classification (VanderWeele-Robins / Weinberg)

06 Linear Gaussian SEM simulation

07 Crude vs adjusted OLS bias computation

08 Canonical example lookup (T01–T15, EM01–EM20)

09 Engine self-validation. Callable concordance check against the Pearl 2009 and dagitty reference battery, for agents that need to verify engine state before use.

Status

Source public in the same repository as the editor (dag-studio-mcp/ subdirectory).
npm publication pending. Once shipped, one-line install in any MCP-compatible client.

Install (Coming soon)

One-line install will be available across Claude Desktop, Claude Code, Cursor, and Cline at npm release. Notify me →

Validation & Scope

Validated against two reference implementations.

DAG Studio's identification engine is gated against two reference implementations: the canonical DAGs and identifiability results in Pearl (2009), and the dagitty R package. The Validation tab in the live editor reports the current concordance: 15/15 across the canonical battery. Open the Validation tab →

What this does not do

Not a substitute for methodological judgment about DAG specification.
Does not generate DAGs from raw data. The analyst supplies the candidate DAG.
Does not perform statistical estimation beyond illustrative SEM simulation.
Identification results are only as good as the underlying causal model. The engine reasons over the DAG you provide; it does not validate the DAG itself.

License

MIT

Citation

Preprint forthcoming. Citation block will be updated on release.

Get Involved

Open source by design.

DAG Studio is open source. Issues and pull requests are welcome on the GitHub repository. Inspectable tooling is the right shape for software that may inform regulatory submissions, and community review makes the engine better.

If your team is considering integrating DAG Studio into protocol development workflows, a small pilot program is open. Pilot scope: two protocols, four weeks, written feedback in exchange for early access and direct engine support. Pharmacoepidemiology and real-world evidence teams are the primary audience.

Get in touch →