Alchemy · Engineering

How Alchemy runs

An AI-native operating layer, built on a knowledge base that already exists.
Capabilities live
13
Knowledge pages
71
Cross-references
1,329
Human sign-off
Always
Most companies bolt AI onto their work as an afterthought — a chat window beside the real system. Alchemy inverted that. The institutional knowledge came first: a structured, cross-referenced knowledge base built from primary engineering and business documents, served to every surface over the Model Context Protocol. The agents came second, and they all read from the same brain.

The knowledge base came first

Every agent below reads from one source: a structured knowledge base of interlinked pages, generated from primary source documents — protocol specifications, architecture records, technical designs — rather than from anyone's recollection. It is served over MCP with public and internal views separated at the server, so an agent talking to the public physically cannot reach internal material. The knowledge base reports its own size at a public health endpoint.

The map

Every capability branches from the same knowledge base. Drag to pan, scroll to zoom, and select any node to read what it does.

drag to pan · scroll to zoom · tap a node
Knowledge baseContinuous ingestFreshness monitoringFOUFoundationSite conciergeMeeting conciergePUBPublic surfacesRetrieval-grounded answeringCitation enforcementANSAnswer qualityAutomated language reviewHuman sign-off by designSensitivity boundariesRegulatory monitoringGOVGovernanceProtocol code reviewInvariant and fuzz testingAudit readinessOn-chain monitoringOracle integrity checksENGEngineeringBuilder supportEcosystem monitoringECOEcosystem Knowledgebase

Foundation

Knowledge baselive

A structured, interlinked knowledge base generated from primary source documents and served over MCP. Public and internal views are separated server-side.

Continuous ingestlive

New source documents are read, cross-referenced and written into the knowledge base, so the brain stays current as the protocol ships.

Freshness monitoringlive

Automated checks flag knowledge that has aged relative to what has shipped, so public-facing agents deflect rather than answer from stale material.

Public surfaces

Site conciergelive

Answers ecosystem questions on alchemy.xyz from the public knowledge base, with retrieval-grounded answers and a refusal path when the knowledge base does not cover the question. try it →

Meeting conciergelive

A conversational scheduling agent that handles timezone-aware booking end to end.

Answer quality

Retrieval-grounded answeringlive

Answers must trace to a source page. A relevance gate rejects near-miss matches — if the knowledge base does not genuinely cover a question, the agent says so instead of assembling a plausible non-answer from adjacent pages.

Citation enforcementlive

Every generated answer carries the source pages it was built from, so any statement can be traced back to the document behind it.

Governance

Automated language reviewlive

Outbound text is checked against the communication standards a regulated, listing-track business is held to — including hedging on forward-looking statements. Drafts that fail are withheld for human rewrite, never auto-corrected.

Human sign-off by designlive

No agent sends external communications, moves value, or commits the company. Agents draft and analyse; people decide and send.

Sensitivity boundarieslive

Categories of question are refused by policy regardless of what retrieval returns, so the boundary does not depend on a model choosing well in the moment.

Regulatory monitoringlive

Weekly scan of policy movement across the jurisdictions that matter for gold-backed tokens and stablecoins.

Engineering

Protocol code reviewlive

Every pull request across the protocol repositories is screened for the failure classes that matter in financial smart contracts — unit and decimal handling, cross-contract address assumptions, and test coverage moving with the code.

Invariant and fuzz testingbuilding

Continuous property-based testing across protocol repositories, with automatic triage of new failures.

Audit readinessbuilding

Assembles the specification-to-code map and known-issue register that external auditors work from.

On-chain monitoringbuilding

Watches protocol-controlled contracts and treasury movements, surfacing pending activity before it executes.

Oracle integrity checksbuilding

Cross-checks the gold price feed against independent references and alerts on divergence, staleness, or halt.

Ecosystem

Builder supportbuilding

Onboards teams building on the network and answers technical questions from the public knowledge base.

Ecosystem monitoringlive

Tracks the tokenised-gold and stablecoin landscape — launches, volumes, and incidents — so positioning stays grounded in what the market is actually doing.