AURORA Agent

Context assembly as a compiler pass

Context engineering, with receipts.

AURORA Agent is a local Model Context Protocol server built on FIBER, a decision context compiler. Instead of retrieval plus summarisation plus ad-hoc compaction, a typed decision query is compiled into the smallest decision-sufficient evidence region — delivered as a Decision Section, with a Context Certificate that states exactly what was omitted and whether the omission could have changed the decision.

It is a Rust workspace of 88 crates (engine crates are prefixed bioprism-), exposing 983 MCP tools over stdio, a CLI (bioprism --json, with a ten-exit-code retryability matrix), and an optional local HTTP gateway (bioprism-api). Apache-2.0, source on GitHub.

What the measurement says

The shipped reference world contains 761 facts. FIBER compiles the query down to 11 facts — 1.45% of the world — and the deterministic oracle still returns the correct verdict with all four leakage witnesses intact.

The headline benchmark is a deliberate negative. On that same reference world, a 5-hop graph walk and a BM25 retriever select the identical eleven facts. FIBER has no measurable advantage there; it separates from both baselines only on a generated world built to vary structure independently. Publishing the tie is the methodology: a compression claim without equal-engineering baselines is unfalsifiable. Details on the benchmarks page and in docs/FINDINGS.md.

Correctness is pinned by cross-implementation parity: the certificate digest c0da17ffc80465258345c8a538171bfd868100cd883e9a20780a0dc5477e7ea4 is agreed byte-for-byte by three implementations — the CPython reference, the eager Rust path, and the Rust indexed store.

What you get as an MCP client

Autonomous workflows, with receipts

bioprism autopilot drives an instantiated workflow's mission autonomously under an explicit AutonomyGrant — the only source of authority; there is no default grant. Every failed step is classified by its declared retry class, and only what the grant authorises is re-dispatched as a repair subset; success is never inferred — it requires full step coverage, a succeeded mission report, and (by default) a complete workflow reconciliation with valid integrity. Every drive emits a digest-sealed report that bioprism autopilot verify recomputes, detecting a single tampered byte.

What it deliberately does not do: no scheduling or recurrence, no MCP tool exposure of the driver itself, no cross-process resume, no wall-clock deadlines. See the FAQ entry and docs/AUTOPILOT.md.

Autonomous research, with receipts

bioprism research template | run | verify executes a fixed protocol over synthetic decision worlds — generate, compile and certify, equal-engineering baseline panel, structural sweep, metamorphic mutation, minimize — and writes a digest-sealed dossier in which every finding cites the sha256 of the artifact supporting it. The committed worked example runs 12 steps in about four seconds and returns nine findings, seven of them negative: FIBER is tied by directed-walk-full at every distractor level and is not separated in 36 of 36 sweep cells.

Measurement over synthetic worlds only — no biology, literature coverage, or external-world claims, and oracle review stays a human gate. See the research page and docs/RESEARCH.md.

Install

Claude Desktop
Download aurora-agent.mcpb (v0.1.1, prebuilt for Windows only) from the GitHub releases and double-click it.
Claude Code
claude plugin marketplace add AURORA-NEURO/aurora-agent then claude plugin install aurora-agent@aurora
MCP registry
Listed as io.github.MurariAmbati/aurora-agent v0.1.1 on registry.modelcontextprotocol.io.
Any MCP client
Build from source with Cargo and run bioprism-mcp over stdio — see the install page.
Research and developer infrastructure: it does not diagnose an individual, recommend treatment, triage care, enroll participants, or claim medical-device functionality.