Frequently asked questions
What is AURORA Agent?
AURORA Agent is a local Model Context Protocol (MCP) server built on FIBER, a decision
context compiler. A typed decision query is compiled into the smallest decision-sufficient
evidence region, delivered as a Decision Section plus a Context Certificate stating exactly
what was omitted and whether the omission could have changed the decision. It is a Rust
workspace of 88 crates (engine crates prefixed bioprism-) exposing 983 MCP tools,
a CLI, and a local HTTP gateway, licensed Apache-2.0.
What is a Context Certificate?
A machine-readable receipt that accompanies every compiled Decision Section. It states exactly which facts were omitted and whether each omission could have changed the decision — distinguishing zero influence from unknown influence. Certificates hash canonical bytes, so a consumer (an MCP client, a CI gate, an auditor) can read and verify a compiled context without linking the engine that produced it.
What does “context engineering, with receipts” mean?
Context engineering is the practice of selecting what an LLM sees. AURORA Agent adds a
receipt: the Context Certificate, which states what was omitted from the compiled context and
whether the omission could have changed the decision. The receipt is machine-checkable — the
fiber_verify tool (or bioprism context verify) verifies a
certificate without linking the engine that produced it.
How is this different from RAG?
Retrieval-augmented generation ranks passages by similarity and hopes the top-k suffice. FIBER computes a mandatory protected closure before any relevance step, slices dependencies from a typed world, and fails rather than truncating when the budget is smaller than the closure. The output carries an explicit sufficiency claim and an omission account, not a similarity ranking. The benchmark findings show why the closure matters: on a world built to discriminate, BM25 reached the correct verdict from an incomplete closure — right by luck, having dropped a protected fact that happened not to matter.
Does it need network access?
Local workflows run offline. There is no telemetry, analytics, crash reporting, or account.
Outbound source HTTP is denied by default and requires an operator-configured exact origin plus
caller-plan opt-in; the selected locator path and query are sent to that host over unencrypted
HTTP. The optional bioprism-api gateway binds to an address you choose — loopback
by default. Full policy:
PRIVACY.md.
Why is the headline benchmark a tie?
On the shipped reference world (761 facts), a 5-hop graph walk, a BM25 retriever at k=11, and FIBER select the identical eleven facts, so that world cannot tell the methods apart. FIBER separates from both baselines only on a generated discriminating world. The tie is published deliberately: without equal-engineering baselines, “FIBER compiles a smaller context” would be unfalsifiable. See the benchmarks page.
What does the MCP registry listing mean?
The server is listed as io.github.MurariAmbati/aurora-agent v0.1.1 on
registry.modelcontextprotocol.io. The
registry is a metadata listing with automated validation — it makes the server discoverable to
registry-consuming tools and aggregators. Installation happens through the
.mcpb bundle, the Claude Code plugin marketplace, or a source build.
What platforms are supported?
The prebuilt Claude Desktop bundle (aurora-agent.mcpb) is Windows-only. On
macOS and Linux, build from source — cargo build --release --offline — and
register the bioprism-mcp binary with any MCP client.
What is the license?
Apache-2.0. The repository AURORA-NEURO/aurora-agent is public on GitHub.
Is this a medical device or clinical tool?
No. Research and developer infrastructure: it does not diagnose an individual, recommend
treatment, triage care, enroll participants, or claim medical-device functionality. The MCP
surface includes a medical_boundary_check tool that admits research use cases and
structurally refuses clinical outputs.
How do I know the implementation is correct?
Byte-level parity with the CPython reference runtime is enforced by test and holds across
three implementations: CPython, the eager Rust path, and the Rust indexed store. All three
agree on the certificate digest
c0da17ffc80465258345c8a538171bfd868100cd883e9a20780a0dc5477e7ea4 for the
reference compilation. The benchmark numbers on this site are likewise asserted by tests in
the repository.
Can I use it without Claude?
Yes. bioprism-mcp speaks standard JSON-RPC 2.0 over stdio and follows the MCP
lifecycle, so any MCP client can use it. The repository also ships a dependency-free Python
SDK, a dependency-free TypeScript SDK, and the bioprism CLI, whose
--json envelopes and ten-exit-code retryability matrix are designed for scripted
callers.
Can AURORA Agent run workflows autonomously?
Yes — under an explicit grant. bioprism autopilot drives an instantiated
workflow's mission autonomously under an AutonomyGrant, the only source of
authority: there is no default grant. The driver dispatches the mission in-process, classifies
every failed step by its declared retry class (terminal, retryable-after-change,
retryable-as-is, or unknown), and re-dispatches only what the grant authorises as a repair
subset with rematerialised bindings — terminal and cancelled steps are never re-dispatched,
and unknown failures are never retried unless the grant explicitly opts in. 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
autopilot report chaining the grant digest, every mission and report digest, and every
reconciliation digest; bioprism autopilot verify recomputes it and detects a
single tampered byte. It deliberately does no scheduling or recurrence, exposes no MCP tool
for the driver itself, cannot resume across processes, and has no wall-clock deadlines. Full
reference:
docs/AUTOPILOT.md.
Can it run its own experiments?
Yes — over synthetic decision worlds only.
bioprism research template | run | verify executes a fixed protocol (generate,
compile and certify, equal-engineering baseline panel, structural sweep, metamorphic mutation,
minimize) and writes a digest-sealed dossier plus a rendered report and figures. Findings are
derived by fixed public rules, locked to level observation — no stronger level is
representable in the type system — and each cites the sha256 of the artifact supporting it;
research verify recomputes the digest and catches a one-byte tamper. The committed
worked example runs 12 steps in about four seconds and produces nine findings,
seven of them negative: FIBER is tied by directed-walk-full at
every declared distractor level, both admissible at 11 facts, and is not separated in 36 of 36
sweep cells. The limits are the point: measurement over synthetic worlds only, no biology, no
literature or prior-work coverage, no external-world claims, oracle review remains a human
gate, the sweep deliberately does not vary decision-defining knobs, and negative findings are
first-class results. See the research page and
docs/RESEARCH.md.