A Powerful Database
for AI
Data Harness and Search Tools for your AI
Give your AI the right data, not all the data. CameoDB loads your data into one engine that stores it, indexes every field and controls who can read what, so your AI answers from records that are fresh and relevant.
Your index is always
a little bit wrong
The classic stack writes to the database, then updates a search index on the side. Between those two writes, search disagrees with the data, and an agent reading in that window gets an answer that was never true.
CameoDB has one write path. The search index is a view of the store, so it catches up with every commit and after every restart.
One write path,
one source of truth
Redb for key-value durability, a flexible document model and Tantivy for full-text search, behind one write path in one process.
- Point reads in about 0.1 ms A lookup by id is a single Redb B-tree read, ready the moment the write returns.
- ACID from the first write Redb commits every write durably and carries it safely through a restart.
- Faults stay small A panic fails one request, not the process, and a writer that stops is restarted over the same store.
-
01
Sequence & WAL Redb A monotonic sequence id, logged in the same transaction as the document.
-
02
KV insert Redb The document body lands in the data table, for O(log N) retrieval by id.
-
03
Indexing Tantivy Fields are parsed and handed to the index writer for full-text search.
-
04
Dual commit RedbTantivy Redb commits with fsync; Tantivy commits on a steady clock and records how far it has read, so recovery picks up right where it left off.
Point your agent at it.
That is the integration
CameoDB speaks MCP natively, on the same port as its HTTP API. Claude Code, Claude Desktop, Cursor and Windsurf connect with one line, and ChatGPT and claude.ai connect by URL with OAuth.
- Indexes that describe themselves Every field’s type, tokenizer and operators, read straight from the engine, so an agent composes queries from what the data holds.
- Scoped to its key An agent is offered only the tools and indexes its key may use, and every call is checked against both.
- Every answer explains itself Shards answered, the schema version searched, and any clause set aside, by name.
# Claude Code
claude mcp add --transport http cameodb \
http://localhost:9480/mcp
Entries for Claude Desktop, Cursor, Windsurf and OAuth clients are on the Build page →
Only what the model
asked for
Stop paying tokens for fields the model never needed. CameoDB projects inside the engine, on the shard that holds the document, so only the fields you request are serialized and sent.
- Hits sized for their reader A field of a hundred thousand values comes back trimmed, matching elements first and their positions named, so the evidence for a hit is in the hit.
- Each value stored once A ghost field searches one stored text through as many language analyzers as it needs, with no second copy.
fields: [title, author, publication_date]
- title
- Animal Farm
- author
- George Orwell
- publication_date
- 1945-08-17
Every column is
a searchable column
One syntax across key-value, document and full-text data: expressive enough for operators, predictable enough for agents to generate.
Language-aware in ten languages
A search for one form of a word finds the others, with or without accents. Croatian runs on a lemmatizer of our own: 1.7 million word forms, 97% F1 on hand-checked text.
field:valueExact term match on a field.field:quick*Term prefix: finds quickstart.field:IN [ab cd ef]Term set, quick on large value lists.field:"looks good to me"Exact phrase across a positional field.body:"looks good"~2Up to N extra words between the terms.ip:[127.0.0.1 TO 127.0.0.50]Inclusive range over IPs, numbers or dates.date:>=2024-01-01Open-ended bound with natural ordering.No leader. No quorum.
No coordinator to lose
Nodes join and leave without an election. Every node can route any request, and the mesh answers as one.
- Consistent hashing Every node computes a key’s owner from the same ring, so placement needs no coordination.
- Kademlia DHT Point a node at one seed and it learns the rest of the mesh from its peers.
- Scatter-gather Queries fan out in parallel and merge on return, so no node becomes the bottleneck.
Production posture
not a promise
End-to-end figures from cameodb-bench,
the harness in the repository, observed by the client over HTTP. Every
release ships with the artifacts an enterprise review asks for.
- Verifiable, not just claimed Every artifact is cosign-signed with a published SHA-256, with SPDX and CycloneDX SBOMs. Check a build yourself
- Open licensing An Apache-2.0 core and MIT clients; the server is FSL-1.1-Apache-2.0 and becomes Apache-2.0 two years after release.
wal_sync off.Closed-loop service times at a fixed concurrency, on a 15-core Apple M5 Pro and an eight-core Linux node, with settings tuned for throughput.
Download, run, query
Six commands from an empty directory to a query answered. The MCP endpoint comes up with the server, so an agent can ask the same question right away.
# macOS: grab the binary curl -sSO https://dl.cameodb.com/mac/cameodb chmod +x ./cameodb # start the server on :9480 ./cameodb # in a second terminal, open the client ./cameodb client -i # load a public dataset, schema inferred cameodb@localhost ▶ data load books https://dl.cameodb.com/examples/data/booksummaries.tsv # search what you just indexed cameodb@localhost ▶ search books title:"Harry Potter" limit 7