Neon Postgres

FIRST: Use the parent neon skill for a Neon overview, getting started with Neon, Neon development best practices, and more.

If the neon skill is not installed, fetch it from https://neon.com/docs/ai/skills/neon/SKILL.md (opens in a new tab) or install it with:

neon skills -s neon -y

Lakebase Postgres

Lakebase Postgres is the database at the core of Neon. It runs on the lakebase architecture — OLTP built directly on cloud object storage — which decouples storage from compute to offer autoscaling, branching, instant restore, and scale-to-zero. It’s fully compatible with Postgres and works with any language, framework, or ORM that supports Postgres.

It is the same database whether you reach it through Neon or through Databricks; this skill covers the Neon access path.

Login, users, sessions, and @neondatabase/auth belong in neon-auth.

Setup Flow

1. Select the organization and project

If a DATABASE_URL is already supplied (prompt, environment, or repo) or a .neon file points at a project, use it. Do not list organizations or create a second project for schema work.

Otherwise use the CLI (default) or MCP server to list organizations and projects. Let the user select an existing project or create a new one.

2. Get the connection string

If a DATABASE_URL is already supplied, use it. Do not fetch another through the CLI or MCP.

Otherwise use the CLI (default), neon env pull, or the MCP server to get the connection string. Store it in .env as DATABASE_URL. Read the file first before modifying it, to avoid overwriting existing values.

When to use pooled vs direct connections

3. Pick the connection method and driver

Preserve the existing ORM and driver. For new TypeScript schema work with no established choice, Drizzle is a suggestion: https://neon.com/docs/guides/drizzle.md (opens in a new tab). Refer to the connection methods guide to pick the correct driver based on how the runtime treats your code: https://neon.com/docs/connect/choose-connection.md (opens in a new tab).

Driver notes:

• On Vercel, use node-postgres (npm install pg) with Vercel Fluid compute and import { attachDatabasePool } from "@vercel/functions";
• On Cloudflare, use node-postgres with Cloudflare Hyperdrive
• On Neon Functions, use node-postgres, as the functions are long-running and reuse the pool across requests.
• Use the @neondatabase/serverless driver for serverless and edge environments (for example, when using Netlify) — HTTP transport for one-shot queries, WebSocket for transaction support. Link: https://neon.com/docs/serverless/serverless-driver.md (opens in a new tab)

4. Set up the schema

Manage schemas and migrations as code. Avoid running ad hoc schema migrations against your database, since they’re hard to manage.

If you’re using an ORM, follow your ORM’s best practices to manage schemas and migrations. For example, if using Drizzle, only use Drizzle for schema and migration management unless instructed otherwise.

Branching

Use this when the user is planning isolated environments, schema migration testing, preview deployments, or branch lifecycle automation.

Key points:

• Branches are instant, copy-on-write clones (no full data copy).
• Each branch has its own compute endpoint.
• Use the neon CLI or MCP server to create, inspect, and compare branches.

Link: https://neon.com/docs/introduction/branching.md (opens in a new tab)

For detailed branch creation workflows (normal vs schema-only branches, reset-from-parent, CLI/MCP selection), use the neon-postgres-branches skill. If it isn’t installed, fetch it from https://neon.com/docs/ai/skills/neon-postgres-branches/SKILL.md (opens in a new tab) or install it with:

neon skills -s neon-postgres-branches -y

Migrations

Test a migration on a branch of production, against production-like data, before applying it to production.

Use a direct (non-pooled) connection string when you run the migration, not a pooled one. neon connection-string returns the direct string by default; make sure the hostname does not include the -pooler suffix.

Troubleshooting and Neon-Specific Performance

Use Neon’s predefined, read-only diagnostics before writing catalog queries by hand. The Neon CLI neon inspect db subcommands and the Neon MCP server’s inspect_database tool run the same checks.

This section covers Neon-specific diagnostic tools, compute cache behavior, and platform signals. When the evidence points to generic Postgres work such as rewriting a query, choosing an index, changing a schema, or interpreting plan nodes, load the postgres-best-practices (opens in a new tab) skill and carry the diagnostic evidence into that workflow.

Docs:

• CLI: https://neon.com/docs/cli/inspect.md (opens in a new tab)
• Query performance: https://neon.com/docs/postgresql/query-performance.md (opens in a new tab)
• pg_stat_statements: https://neon.com/docs/extensions/pg_stat_statements.md (opens in a new tab)
• Neon Local File Cache: https://neon.com/docs/extensions/neon.md (opens in a new tab)

Choose CLI or MCP

Prefer the Neon CLI when terminal access and authentication are available:

neon inspect db <check>

The CLI resolves the project and branch from the current Neon context. Use --project-id, --branch, and --database-name to override it. Omit --database-name to inspect every database on the branch. Use --db-url only when inspecting a Postgres database directly instead of resolving it through the Neon API.

When using Neon MCP, call inspect_database with projectId and one check. Pass branchId, databaseName, or computeId only when needed. Omit databaseName to inspect all databases on the branch. Increase limit only when the result says it was truncated.

Pick the Diagnostic

Do not confuse these checks:

• long-running-queries reports statements running right now for more than five minutes.
• outliers ranks the top queries by cumulative execution time since statistics were reset. It does not rank by mean latency.
• calls ranks by execution count over the same statistics history.

outliers and calls require pg_stat_statements. lfc-hit-rate and working-set require the neon extension. If a check reports a missing extension, ask before running the suggested CREATE EXTENSION statement because installing an extension modifies the database.

Interpret Results Safely

• Treat unused-indexes as a candidate list, not permission to drop indexes. Confirm the observation window, constraints, and workload before removal.
• A sequential scan can be correct for a small table or a query reading much of a table. Check table size, selectivity, and the query plan before adding an index.
• bloat is a statistical estimate. Confirm the impact and plan locks or maintenance before VACUUM FULL, REINDEX, or similar remediation.
• Cache and Postgres statistics reset when compute restarts, including scale-to-zero suspension. Run a representative workload before interpreting fresh lfc-hit-rate, working-set, vacuum-stats, or pg_stat_statements results.
• Compute-wide checks (lfc-hit-rate, working-set, and replication-slots) run once even when inspecting every database.
• One failing database can fail an all-databases inspection; retry the relevant check with an explicit databaseName to isolate it.

Inspect Neon Cache Behavior Per Query

Standard EXPLAIN (ANALYZE, BUFFERS) reports Postgres shared-buffer activity, but it does not show Neon’s Local File Cache (LFC) or page prefetching. For a safe read-only query, add Neon’s FILECACHE and PREFETCH options:

EXPLAIN (ANALYZE, BUFFERS, PREFETCH, FILECACHE)
SELECT ...;

• File cache: hits counts pages found in the compute’s LFC.
• File cache: misses counts pages not found in the LFC and fetched from database storage.
• Prefetch: hits, misses, expired, and duplicates show how effectively Neon fetched pages before the executor requested them.

FILECACHE and PREFETCH provide metrics for this query and do not require the neon extension. By contrast, neon inspect db lfc-hit-rate and working-set provide compute-wide statistics and do require the extension.

The MCP explain_sql_statement tool can produce a standard plan but does not expose FILECACHE or PREFETCH options. To collect those Neon-specific metrics through MCP, use run_sql with the explicit, read-only EXPLAIN statement above.

Because ANALYZE executes the statement, use it only when execution is safe; do not run it autonomously for mutating SQL. Compare cold- and warm-cache runs carefully because the first execution can populate the cache and materially change later results.

Performance Workflow

1. Reproduce the symptom and note its time window.
2. Run the smallest relevant inspect checks from the table above.
3. Identify a specific query before changing schema or compute. Use MCP explain_sql_statement for a standard plan, or the Neon-specific EXPLAIN above when LFC or prefetch behavior matters.
4. If the bottleneck is query shape, indexing, schema, locking, or vacuum behavior, load postgres-best-practices and carry forward the inspection results and query plan. Keep Neon compute, cache, connection, and platform decisions in this skill.
5. Re-run the same check and workload to verify the change.

Use MCP list_slow_queries instead of inspect_database when the user specifically needs queries ranked by average execution time with a custom threshold and limit. Outside the explicit EXPLAIN case above, use run_sql only for read-only diagnostic SQL when the predefined checks do not answer the question.

Autoscaling

Use this when the user needs compute to scale automatically with workload and wants guidance on CU sizing and runtime behavior.

Link: https://neon.com/docs/introduction/autoscaling.md (opens in a new tab)

Scale to Zero

Use this when optimizing idle costs and discussing suspend/resume behavior, including cold-start trade-offs.

Key points:

• Idle computes suspend automatically after a default of 5 minutes; the timeout is configurable, and suspension can only be disabled on the Launch and Scale plans.
• First query after suspend typically has a cold-start penalty (around hundreds of ms)
• Storage remains active while compute is suspended.

Link: https://neon.com/docs/introduction/scale-to-zero.md (opens in a new tab)

Instant Restore

Use this when the user needs point-in-time recovery or wants to restore data state without traditional backup restore workflows.

Key points:

• History windows for instant restore depend on plan limits.
• Users can create branches from historical points-in-time.
• Time Travel queries can be used for historical inspection workflows.

Link: https://neon.com/docs/introduction/branch-restore.md (opens in a new tab)

Read Replicas

Use this for read-heavy workloads where the user needs dedicated read-only compute without duplicating storage.

Key points:

• Replicas are read-only compute endpoints sharing the same storage.
• Creation is fast and scaling is independent from primary compute.
• Typical use cases: analytics, reporting, and read-heavy APIs.

Link: https://neon.com/docs/introduction/read-replicas.md (opens in a new tab)

Connection Pooling

Use this when the user is in serverless or high-concurrency environments and needs safe, scalable Postgres connection management.

Key points:

• Neon pooling uses PgBouncer.
• Add -pooler to endpoint hostnames to use pooled connections.
• Pooling is especially important in serverless runtimes with bursty concurrency.

Link: https://neon.com/docs/connect/connection-pooling.md (opens in a new tab)

IP Allow Lists

Use this when the user needs to restrict database access by trusted networks, IPs, or CIDR ranges.

Link: https://neon.com/docs/introduction/ip-allow.md (opens in a new tab)

Logical Replication

Use this when integrating CDC pipelines, external Postgres sync, or replication-based data movement.

Key points:

• Neon supports native logical replication workflows.
• Useful for replicating to/from external Postgres systems.

Link: https://neon.com/docs/guides/logical-replication-guide.md (opens in a new tab)

Lakebase Search

Use Lakebase Search for semantic, full-text, and hybrid search:

• For semantic search, read Vector search (opens in a new tab).
• For full-text search with BM25 ranking, read Full-text search (opens in a new tab).
• For combining semantic and lexical results, read Hybrid search (opens in a new tab).
• For managing any of the above through Drizzle ORM, read Managing Lakebase Search with Drizzle (opens in a new tab).

Links:

• Get started with Lakebase Search (opens in a new tab)
• lakebase_vector reference (opens in a new tab)
• lakebase_text reference (opens in a new tab)

Gotchas

Pooled vs direct connections: use the direct URL for migrations, dumps, and replication

Neon gives you two connection strings for the same database: a pooled one (hostname with the -pooler suffix) and a direct/unpooled one (no -pooler suffix). neon env pull writes them as DATABASE_URL and DATABASE_URL_UNPOOLED. The pooled connection routes through PgBouncer in transaction mode, which doesn’t support session-level operations. Choose the right one:

• Pooled (DATABASE_URL) — your application’s normal query traffic, especially serverless and connection-per-request workloads.
• Direct (DATABASE_URL_UNPOOLED) — schema migrations (Prisma Migrate, Drizzle Kit, Alembic, and others), pg_dump / pg_restore, logical replication, LISTEN/NOTIFY, and anything relying on SET or other session state.

Running migrations, dumps, or replication over the pooled connection can fail, and never in a way that names pooling: prepared statement "s0" already exists from Prisma Migrate, a SET search_path that doesn’t persist past its own transaction so the next query reports relation "mytable" does not exist, or a write intermittently hitting a read-only transaction (SQLSTATE 25006) that a pooled backend inherited from an earlier client. Migration tools generally take both strings at once — Prisma’s directUrl alongside url — so point that at the direct one rather than swapping DATABASE_URL and losing pooling for the application. See https://neon.com/docs/connect/connection-pooling.md (opens in a new tab).