LessDB — Product Positioning

LessDB — Product Positioning

Status: adopted · Owner: product · Updated: 2025

How LessDB positions itself — a fit that comes from what the product already is, validated against market research. Evidence base: RESEARCH.md (four compiled streams in docs/research/). Product consequences: AGENT-GOVERNANCE.md.


1. The position (one sentence)

LessDB is the analytical database your agents and your analysts share. One governed SQL system of record — MCP-native for agents, built for analysts — with identical governance on both sides.

Supporting line (the two-doorway frame):

One engine, two front doors. Agents work it over 27 MCP tools; humans work it over SQL, the REPL, and the SDKs. Same engine, same data, same control, one audit trail.

2. Why this is a natural fit, not a forced one

The product already is this in three verifiable ways:

  1. Two doors already exist. less sql (human) and less mcp
  2. (agent) are two commands on one binary over the same tables. No narrative contortion required — the demo shows it.

  1. The multi-model store already exists. SQL tables, vector spaces,
  2. a linked knowledge graph, and RAM tables share one data dir. Research shows the market forces an either/or choice (vector DB or graph DB or SQL store) while agents actually juggle all three — we remove the forced choice (RESEARCH §1, gaps).

  1. The governance gap is the market's loudest pain. ~200k exposed MCP
  2. instances, "the protocol that trusted everyone", agents inheriting full human DB roles, and a viral incident where a coding agent wiped a startup's database and backups in 9 seconds (RESEARCH §1–§3). Our answer — one role model, per-agent identity, one audit trail — is the feature set enterprise buyers already articulate as their "full control" checklist: per-agent identity, row/column permissions, immutable audit, write-approval gates (RESEARCH §2).

3. What we are NOT (the discipline)

4. Positioning pillars (messaging hierarchy)

  1. One database, two front doors. Agents over MCP, humans over SQL —
  2. same engine, same data, no sync, no exports, no drift.

  1. Full control and governance for both. One identity model, one role
  2. model, one audit trail. An agent is just another caller — with a credential, a role, and a record of everything it did. (The wedge: agent data stores lack audit; governed databases lack agents.)

  1. More with less. Columnar storage with zstd, bloom pruning, 42 MB
  2. RSS, GPU kernels, stateless compute — the resource-efficiency story the name promises.

  1. Boring, open, embeddable. Rust, Apache-2.0, one binary, Arrow-native.
  2. Works with the tools both audiences already use.

5. Competitive frame (from research)

NeighborOwnsDoesn't ownOur wedge
Vector DBs (hosted/embedded)Embedding searchSQL, governance, graphThe store vectors graduate into when they must meet tables and audits
Embedded analytics enginesHuman SQL, no serverFirst-class agent door with permissionsSame power plus a governed, MCP-native agent door
Memory frameworksSession state as a library/appThe store underneath (their moat problem)We are the store underneath — not the app on top
Warehouses / lakehouse platformsEnterprise governance + scaleEmbeddability, open single-binary, agent-native toolingThe embedded, open-source version of that trust
MCP gateways (Kong/Tyk/Gravitee…)Enforcing policy in front of other databasesBeing the databaseGovernance built into the store itself — one less hop, one audit trail

6. Audiences (the two doorways)

AgentsHumans
DoorMCP (stdio), 27 toolsCLI/REPL, Python/Node SDKs, HTTP API
WorkQuery, search, link, persist context across sessionsSQL analytics, ingestion, dashboards, pipelines
NeedsTool-shaped access, cheap calls, persistence, permission to actFull SQL, ecosystem interop (Arrow/pandas), visibility + control
SharedThe same tables, vectors, graph — and the same governance

The doorway pattern follows the DuckDB model (one architectural truth, audience-specific entrances) — not two products (RESEARCH §4).

7. GTM shape

8. Website changes implied

9. Product changes implied (priority order)

See AGENT-GOVERNANCE.md for the full spec.

  1. P0 — close the trust gap: MCP auth (SEP-1046-style client
  2. credentials), tool→permission map, append-only audit log.

  1. P0 — table-level grants in the catalog (cheap, unblocks tenancy
  2. completion).

  1. P1 — human approval gates (propose→approve→commit for agent writes)
    • quotas + schema-governance flags.
  1. P1 — standards alignment: MCP tool annotations surfaced as
  2. permissions; OAuth 2.1/PKCE on the MCP door; OPA/Cedar policy hook.

  1. P2 — the enterprise test: row-level security + column masking, then
  2. data lineage (OpenLineage) — the two primitives buyers check first.