Dolt as Agentic Database¶
Type: Database technology / architectural enabler
Definition¶
Dolt is a SQL database that combines the functionality of Git with the query power of MySQL/Postgres. It provides branches, commits, diffs, merges, and conflict detection — all accessible through standard SQL. DoltHub positions it as "the database for agents" because it provides the version control primitives that make LLM writes safe and reviewable.
How It Works¶
Dolt is a drop-in replacement for MySQL or Postgres:
- MySQL flavor: github.com/dolthub/dolt
- Postgres flavor: github.com/dolthub/doltgresql
Migration: dump your existing database, import into Dolt, change the connection string. Your application works exactly as before.
Version control is exposed through SQL:
- Branches — CALL dolt_branch('my-branch')
- Commits — CALL dolt_commit('-am', 'message')
- Diffs — query dolt_diff() like a table
- Merges — CALL dolt_merge('branch-name')
- Conflict detection — automatic on merge
Key Properties¶
- Drop-in compatible — MySQL/Postgres wire protocol
- Git-like version control — branches, commits, diffs, merges in SQL
- Queryable diffs — diffs are tables you can JOIN and filter
- Conflict detection — prevents overwrites when concurrent changes exist
- Full history — permanent audit trail of all changes
- SQL procedures — all operations available through standard SQL
Why It Matters for Agentic Applications¶
Without version control in the database:
- LLM writes are invisible (no diff)
- LLM writes are irreversible (no rollback)
- LLM writes are dangerous (can overwrite real data)
With Dolt:
- Every LLM write produces a diff
- Rejection is trivial (delete branch)
- Conflict detection prevents data loss
- Full audit trail of agent actions
Related Concepts¶
- cursor-for-everything — the pattern that uses Dolt
- branch-aware-llm-writes — the specific write pattern enabled by Dolt
- versioning/git-automation — Git-level version control (different scope)
- mcp-model-context-protocol — how LLMs connect to Dolt-backed apps
Source¶
- cursor-for-everything — full source article