Biolab MCP Server

"AI agents querying biological databases leave no audit trail. Six months later, nobody can answer: what exact query returned this result, when, and was that paper peer-reviewed at the time? Biolab solves that."

A Python MCP server that sits between AI agents and biological databases. Every query is intercepted, logged with full retrieval context, and returns a retrieval_id that calling systems can store alongside their own reasoning traces — creating an end-to-end auditable chain from conclusion back to raw source.

---

The Problem

A drug discovery team uses an AI agent to research gene targets. The agent queries PubMed 200 times over three days and surfaces a paper claiming gene X is upregulated in pancreatic cancer. A scientist makes a decision based on that. Six months later, during FDA submission:

• What exact query returned that paper?
• What date was it retrieved?
• Was it peer-reviewed at retrieval time, or a preprint that was published later?
• Did the agent summarize it accurately, or did it hallucinate details?

Without Biolab, nobody can answer any of those questions. The retrieval is invisible.

---

What Biolab Does

Biolab is an interception and logging layer, not a retrieval layer.

When an agent calls the PubMed tool:

Aletheia Advocate Agent
    ↓  MCP tool call
Biolab MCP Server
    ↓  HTTP
PubMed API
    ↓  paper
Biolab writes retrieval record to database
    ↓  paper + retrieval_id
Back to Advocate Agent

The agent gets the paper it asked for. Biolab gets a permanent, queryable record of exactly what happened.

---

How It Fits Into Aletheia

Biolab is a dependency of Aletheia, a multi-agent scientific reasoning system. Aletheia's advocate agent retrieves evidence through Biolab. Aletheia stores retrieval_id alongside source_paper_id in its own provenance table.

This creates the link:

Aletheia provenance table
    claim | agent | source_paper_id | retrieval_id | action | timestamp
                                          ↓
                              Biolab retrieval record
                              (full query context, abstract at retrieval time, evidence level)

Without retrieval_id, you know which paper was used but not what it said when it was retrieved. With it, the chain is complete.

---

Why Every Decision Exists

Python, not Go

The official MCP SDK is Python. Aletheia is Python. The biotech ecosystem is Python. A Go service introduces a language boundary at the most critical integration point with no performance justification — Aletheia makes sequential agent calls, not 10,000 concurrent ones. Python eliminates the boundary entirely.

MCP tool, not REST API

Aletheia's agents call tools, not endpoints. Wrapping Biolab as an MCP tool means zero integration overhead on the Aletheia side — the agent calls it exactly like any other tool in its environment.

Database, not log files

An audit trail needs to be queryable. "Show me everything retrieved for BRCA1 between June and August" is a SQL query, not a grep. Log files cannot answer structured questions across time. A database can.

No LangChain, no LangGraph

Same reason as Aletheia: these frameworks hide the exact artifact at each step inside abstractions you don't control. Provenance tracing is the core product. Every step must produce an inspectable record. That requires code you own.

---

Retrieval Log Schema

Status: In design. The starter schema is below. The open question is what additional columns are needed to fully satisfy a regulatory audit — specifically around paper status at retrieval time.

retrieval_id    → UUID, primary key
query_text      → the exact search string sent to PubMed
pmid            → PubMed paper ID returned
retrieved_at    → UTC timestamp of retrieval
agent_id        → which agent made the call (e.g. "aletheia:advocate")

Pending columns: paper publication status at retrieval time, abstract snapshot, evidence level classification.

---

Stack

Layer	Choice	Why
Language	Python	Official MCP SDK, zero boundary with Aletheia, biotech reads Python
Protocol	MCP	Aletheia agents call tools, not endpoints
External API	PubMed E-utilities	Real, citable, stable biological literature source
Storage	TBD (SQLite → PostgreSQL)	Start simple, migrate when query patterns are known

---

90-Day Deliverable

One live MCP tool: search_pubmed. Takes a query string, retrieves papers from PubMed, writes a retrieval record to the database, returns papers plus retrieval_id to the calling agent.

A demo showing Aletheia's advocate agent calling search_pubmed, receiving a sourced result, and storing the retrieval_id in Aletheia's provenance table — making the full chain from conclusion to raw source queryable.

---

Commandments

1. Don't add infrastructure until a real query fails without it.
2. Every retrieval produces a permanent, queryable record.
3. The retrieval_id is not optional — it is the link that makes Aletheia's traces auditable.
4. Never log the agent's summary instead of the raw source. One is interpretation. One is ground truth.