OpenInterpretability/openinterp-mcp

OpenInterpretability/openinterp-mcp

Skill Claude CodeCodex

Capture residual-stream activations at specified layers and token positions during a forward pass on the attached model. Returns a captureid for downstream tools.

OpenInterpretability/openinterp-mcp

Skill Claude CodeCodex

Run the four mandatory checks (random-feature baseline, control-token norm, structural-rigidity α-sweep, and the paper-11 structure-matched control + naming gate) and return a verdict in {causal, weak-causal, confounded-structural, epiphenomenal-softmax, epiphenomenal-template, inconclusive-unverified, undetermined}.

OpenInterpretability/openinterp-mcp

Skill Claude CodeCodex

Attach to a researcher's running Colab/vast.ai openinterp session via its public HTTPS URL. Run this once at session start before any other openinterp skill.

OpenInterpretability/openinterp-mcp

Skill Claude CodeCodex

Report the health of the active session — loaded model, probes, and captures in memory — to confirm the backend is alive before a run and to debug a broken attach.

OpenInterpretability/openinterp-mcp

Skill Claude CodeCodex

List the probes currently loaded in the Colab backend — their model, layer, position, and source — so you know what is available to evaluate or steer.

OpenInterpretability/openinterp-mcp

Skill Claude CodeCodex

Apply a loaded linear probe to a stored capture. Returns per-sample scores and AUROC (when labels are provided).

OpenInterpretability/openinterp-mcp

Skill Claude CodeCodex

Decompose a captured activation into its top-K sparse-autoencoder features and read their auto-interp names — the bridge from a residual-stream vector to human-readable concepts.

OpenInterpretability/openinterp-mcp

Skill Claude CodeCodex

Inject a directionalpha into the residual stream during a forward pass and observe the behavioral effect. Returns both raw and control-token-normalized deltas.