corpusBLOCK
Semantic search and structured query over your own content
Overview
From the repository's own README, as read at the audited commit. Badges and raw HTML are left out.
Semantic search and structured query over your own content.
corpus is a knowledge base for your own content. Point it at markdown, documents, or structured records — corpus classifies each item, embeds it through an OpenAI-compatible endpoint, and stores the vectors alongside structured metadata in Postgres/pgvector. Then query it by meaning or by fact.
Everything starts from a raw markdown layer that keeps each item verbatim. From it, corpus derives a configurable set of storage tiers: each tier declares its projection, the tool that exposes it, and an access rule naming who may read it. A sensitive tier serves raw plus enrichment to callers you trust with the originals; a one-way sanitize projects only cloud-safe fields into a tier a lower-trust caller — a cloud-model agent — may query. Add as many tiers as your trust boundaries need; access is a property of each tier, not bolted on afterward.
The embedding endpoint is any OpenAI-compatible API. Run a local model and nothing leaves your network; use a hosted provider and the pi
cb024ec05ea5OBSERVED · 2026-10-11Connect
Built from this server's own package name, version and transport as found in its source — not copied from anyone's documentation, so it cannot drift against a page we do not control.
claude mcp add corpus -- uvx corpus
{
"mcpServers": {
"corpus": {
"command": "uvx",
"args": [
"corpus"
]
}
}
}Exposed tools (6)
6 read · 0 write · 0 destructive.
| Tool | Risk | Description |
|---|---|---|
corpus_get | read | Fetch one indexed document by id: full body content and all metadata. |
corpus_stats | read | Index totals and per-label document counts. |
index_by_domain | read | Recent sanitized items in a life-area domain. |
index_due_soon | read | Time-sensitive items, or items carrying a deadline. |
index_stats | read | Counts of sanitized items by domain and suggested disposition. |
index_summary | read | The sanitized summary + priority signal for one message id. |
Trust audit
BLOCKgrade D · trust 65/100 Do not install this without reading the findings. The audit found something that could harm you or your machine.
| Layer | What it checks | Result |
|---|---|---|
| L0 | Provenance & inventory | PASS |
| L1 | Static analysis of the code | FAIL |
| L2 | Instruction surface (what it tells the agent) | PASS |
| L3 | Class-specific surface | WARN |
| L4 | Behavioural (sandbox) | SKIPPED |
What the source does
- Filesystem
- none-observed
- Network
- declared (7 observation(s))
- Shell
- declared (1 observation(s))
- Dependencies
- pinned
- Secrets in source
- found
Findings (18)
["-----BEGIN OPENSSH PRIVATE KEY-----", *lines, "-----END OPENSSH PRIVATE KEY-----"]
lines = [f"## Enrichment eval ({', '.join(names)})", ""]dsn = f"postgresql://postgres:[email protected]:{port}/ai"
"ghp_0123456789abcdefghijklmnopqrstuvwxyz and "
secret = "-----BEGIN OPENSSH PRIVATE KEY-----\nAAAA\n-----END OPENSSH PRIVATE KEY-----"
"-----BEGIN OPENSSH PRIVATE KEY-----\n"
"-----BEGIN OPENSSH PRIVATE KEY-----\nQUJD\n-----END OPENSSH PRIVATE KEY-----"
block = "-----BEGIN OPENSSH PRIVATE KEY-----\nQUJDREVG\n-----END OPENSSH PRIVATE KEY-----"
streamable-http
SCHEMA_VERSION = hashlib.sha1(msgspec.json.encode(json_schema())).hexdigest()[:12]
return hashlib.sha1(repr(parts).encode()).hexdigest()[:12]
stem = uid if _SAFE_UID.fullmatch(uid or "") else hashlib.sha1(doc_id.encode()).hexdigest()[:24]
monkeypatch.setattr(settings, "openai_api_base", f"http://127.0.0.1:{port}/v1")yield f"http://127.0.0.1:{port}/v1"("http://localhost:8080/v1", True), ("http://127.0.0.1:8000/v1", True),monkeypatch.setattr(settings, "openai_api_base", f"http://127.0.0.1:{port}/v1")monkeypatch.setattr(settings, "openai_api_base", f"http://127.0.0.1:{port}/v1")access revokes when that password is rotated. Configure one fetcher name per
Gates applied: critical_finding, no_behavioural_pass.
cb024ec05ea5full audit observations/trust-audit/mcp-server/ryanjdillon__corpus.json · Report an issue / request a re-scanAudit history
Every audit this server has had. A grade with a past is a grade somebody is still checking.
| Date | Source | Verdict | Grade | Score | Change |
|---|---|---|---|---|---|
| 2026-10-11 | cb024ec05ea5 | BLOCK | D | 65 | first audit |
Questions
What is the corpus MCP server?
Semantic search and structured query over your own content
What tools does corpus expose?
6 in total: 6 read-only, 0 that write, and 0 that can delete or overwrite. Every one is listed on this page with its risk.
Is corpus safe to connect to an agent?
No — not without reading the findings first. The audit graded it D (65/100) and found 2 critical or high issues in the source. Each one is listed on this page with the file and line it is on.
What credentials does corpus need?
No credential environment variables were found in its source, so it appears to need none.
How does corpus run?
It speaks streamable-http, so it runs as a service you connect to over the network. It is published on PyPI as corpus.
How current is this page?
The grade is for one exact copy of the source (cb024ec05ea5), read on 2026-10-11. The repository is watched and re-audited when it changes.