Undercroft 1.6.1 is out

Infrastructure for AI that proves itself.

Sealcroft builds the layer underneath your agents — memory, security, evaluation — designed so every claim it makes can be checked rather than believed. Local-first by default. No telemetry, no API keys, no data leaving your machine.

Explore Undercroft View source
99.4%
LongMemEval R@5 · +MiniLM
38
MCP tools
1048
tests passing
0
bytes phoned home

Projects

One layer at a time.

Each project is independent, source-available, and useful on its own. Together they are the substrate an autonomous system needs before it can be trusted with anything real.

Undercroft

Shipping · v1.6.1

Hardened AI memory in Rust. Your agent's words kept verbatim in encrypted vaults — XChaCha20-Poly1305 at rest, an HMAC-SHA256 tag on every record, and a tamper-evident audit chain that catches an offline edit at unlock. Hybrid semantic and lexical retrieval, a temporal knowledge graph, screened writes and receipted deletion. Runs entirely on your machine with no model to download.

RustMCP serverREST /v1 SQLite + 5 vector backendsPost-quantum exportBUSL-1.1

Security posture ·

In development

Continuous posture assessment for systems that have to answer for themselves — built on the same principle as the rest: findings you can reproduce, not a score you have to take on faith.

AI harness ·

In development

An evaluation and orchestration harness for agentic work — measured on real corpora, reported as measured, favourable or not.

How we build

Verifiable beats trustworthy.

Trust is a claim about intent. Verification is a property of the artifact. We build for the second, because it is the only one that survives being wrong.

Nothing leaves by default

No telemetry, no phone-home, no API key required for core function. Outbound paths exist, but every one is an explicit opt-in — never a silent fallback.

Integrity you can check

Records carry their own tags. Deletions hand back receipts. Audit chains are replayable. If something was altered on disk, the software says so instead of serving it.

Honest about limits

Threat models state what they do not cover. Benchmarks are published as measured. A security feature that overstates itself is a vulnerability with good marketing.

Every record carries an HMAC-SHA256 tag, verified on read before a single byte is returned. hmac-sha256 · 9f3c1a…