Agentic Causal AI
Give agents explicit model operations—not just more prose.
Darkstar can connect an authorized AI client to supported probabilistic and causal model operations. The calculation stays in the Darkstar engine; the agent requests an operation and receives a structured result, job, or refusal it can handle explicitly.
How the request travels
MCP carries the request; Darkstar performs the reasoning.
Model Context Protocol (MCP) is the interface between an authorized agent client and Darkstar. It is transport and orchestration—not another inference engine.
- 1AI agent or client
Frames a question and supplies authorized context.
- 2MCP interface
Validates and routes a supported, typed request.
- 3Darkstar model and execution services
Select and execute the compatible operation in the native reasoning boundary.
- 4Structured result
Returns a completed result, durable job identity, or explicit refusal.
A bounded agent workflow
Make each handoff inspectable.
These MCP tools let an agent discover models, make reviewed changes, run supported reasoning, and retrieve durable results. What a request can do depends on the model family, revision, authorization, and runtime limits.
Find the right model and operation
Give the agent a reliable starting point before it reads, changes, or runs anything.
- Check whether an operation is supportedChecks a requested operation for the authorized account and returns the applicable capability and limits, or an explicit refusal.
darkstar.capabilities.resolve - Find accessible modelsReturns a paginated list of models the account may access, including the stable identities needed for later requests.
darkstar.models.list - Open a modelReturns one authorized model or a requested immutable revision so the agent can inspect the exact model it will use.
darkstar.models.read
Create and track models
Create a durable model, update its details safely, and revisit earlier revisions.
- Create a modelCreates a model in a supported family and returns its stable model identity and first revision.
darkstar.models.create - Update model detailsUpdates model metadata only when the caller supplies the expected current revision, preventing accidental overwrites.
darkstar.models.update - Review model historyReturns the model's immutable revision history so the agent can select or compare an exact version.
darkstar.models.revisions.list
Edit structure and parameters safely
Preview family-specific changes first, then commit them as a new revision.
- Preview a structure changeValidates a proposed topology change and explains its impact without modifying the model.
darkstar.models.structure.preview - Apply a structure changeCommits a validated topology change against the expected revision and creates a new durable revision.
darkstar.models.structure.apply - Preview a parameter changeValidates proposed family-specific parameter values and summarizes the change without saving it.
darkstar.models.parameters.preview - Apply a parameter changeCommits validated parameter values against the expected revision and creates a new durable revision.
darkstar.models.parameters.apply
Ask probabilistic and causal questions
Run an explicit numerical operation against an exact, authorized model revision.
- Run a probabilistic taskRuns a supported inference task, such as a query under evidence, and returns a structured result or explicit refusal.
darkstar.tasks.run - Run a causal analysisRuns a supported causal operation with explicit inputs and returns structured estimates, diagnostics, and provenance.
darkstar.causal.run
Run longer work and retrieve results
Move bounded work out of the interactive request, follow its progress, and collect its output.
- Start a durable jobSubmits admitted asynchronous work and returns a stable job identity the agent can monitor.
darkstar.jobs.submit - Check job progressReturns the current state of an authorized job, including whether it is queued, running, succeeded, failed, or cancelled.
darkstar.jobs.read - Locate a job's resultResolves a completed job to the durable result it produced.
darkstar.jobs.result - Read a stored resultReturns an authorized result and its metadata without rerunning the original operation.
darkstar.results.read - Export a resultPrepares an authorized result in a supported downloadable format for use outside the agent conversation.
darkstar.results.export
Understand what happened
Let people and agents trace authorized activity instead of relying on conversational memory.
- Review the audit trailReturns bounded audit events visible to the account so actions and changes can be traced.
darkstar.audit.list
One question, three causal levels
Association → intervention → counterfactual
An agent can keep the business question conversational while making the numerical operation explicit. Natural language does not supply missing causal assumptions.
- L1Association
“Among observed cases, how does recovery differ with treatment?” The agent selects the authorized revision and requests a supported probabilistic query with
darkstar.tasks.run. - L2Intervention
“What recovery distribution does the declared model imply if treatment is set?” The agent submits an explicit intervention task; observation and do-operator semantics stay distinct.
- L3Counterfactual
“For this observed case, what would the same model imply under the alternative action?” The request preserves factual evidence, the alternative action, target, and model assumptions.
Inspectable execution path
Natural-language intent → agent/LLM → MCP tool → Darkstar model revision → native probabilistic or causal execution → structured result or refusal → agent explanation.
Inspect the causal capability catalogTwo integration paths
Embed capabilities in software—or expose them to agents.
SDKs
Build Darkstar operations into your application or service.
A licensed SDK embeds supported capabilities in software you build and operate. It does not create a Darkstar-managed deployment or include Online subscription rights.
Explore SDK deliveryMCP
Let an authorized agent request supported operations.
MCP exposes supported capabilities through an agent-facing interface. It does not guarantee autonomous decision quality, correct real-world assumptions, or improved outcomes.
Discuss an MCP integration