diff --git a/README.md b/README.md
index 049253c..8a1ccc9 100644
--- a/README.md
+++ b/README.md
@@ -105,3 +105,20 @@ You can also pass any native ClickHouse format name directly.
 |------|-------------|
 | `--help` | Show help text |
 | `--version` | Print version |
+
+## Agent Skill
+
+chcli includes an [Agent Skill](https://agentskills.io) so AI coding agents (Claude Code, Cursor, etc.) can query ClickHouse databases on your behalf.
+
+Install the skill:
+
+```bash
+npx skills add obsessiondb/chcli
+```
+
+This gives your agent the ability to:
+
+- Run SQL queries against ClickHouse using `chcli`
+- Explore database schemas (`SHOW TABLES`, `DESCRIBE TABLE`)
+- Extract data in machine-readable formats (JSON, CSV)
+- Follow best practices like using `LIMIT` and structured output formats
diff --git a/skills/clickhouse-query/SKILL.md b/skills/clickhouse-query/SKILL.md
new file mode 100644
index 0000000..ac6025a
--- /dev/null
+++ b/skills/clickhouse-query/SKILL.md
@@ -0,0 +1,147 @@
+---
+name: clickhouse-query
+description: Query ClickHouse databases using the chcli CLI tool. Use when the user wants to run SQL queries against ClickHouse, explore database schemas, inspect tables, or extract data from ClickHouse.
+metadata:
+  author: obsessiondb
+  version: "1.0"
+compatibility: Requires bun or node (for bunx/npx). Needs network access to a ClickHouse instance.
+allowed-tools: Bash(bunx chcli:*) Bash(npx chcli:*) Bash(chcli:*) Read Write
+---
+
+# chcli — ClickHouse CLI
+
+chcli is a lightweight ClickHouse command-line client. Use it to run SQL queries, explore schemas, and extract data from ClickHouse databases.
+
+## Running chcli
+
+Prefer `bunx` if Bun is available, otherwise use `npx`:
+
+```bash
+bunx chcli -q "SELECT 1"
+npx chcli -q "SELECT 1"
+```
+
+Or install globally:
+
+```bash
+bun install -g chcli
+chcli -q "SELECT 1"
+```
+
+## Connection
+
+Set connection details via environment variables (preferred for agent use) or CLI flags. CLI flags override env vars.
+
+| Flag | Env Var | Default |
+|------|---------|---------|
+| `--host` | `CLICKHOUSE_HOST` | `localhost` |
+| `--port` | `CLICKHOUSE_PORT` | `8123` |
+| `-u, --user` | `CLICKHOUSE_USER` | `default` |
+| `--password` | `CLICKHOUSE_PASSWORD` | *(empty)* |
+| `-d, --database` | `CLICKHOUSE_DATABASE` | `default` |
+| `-s, --secure` | `CLICKHOUSE_SECURE` | `false` |
+
+For agent workflows, prefer setting env vars in a `.env` file (Bun loads `.env` automatically) so every invocation uses the same connection without repeating flags.
+
+See `references/connection.md` for detailed connection examples.
+
+## Query Patterns
+
+**Inline query** (most common for agents):
+
+```bash
+bunx chcli -q "SELECT count() FROM events"
+```
+
+**From a SQL file:**
+
+```bash
+bunx chcli -f query.sql
+```
+
+**Via stdin pipe:**
+
+```bash
+echo "SELECT 1" | bunx chcli
+```
+
+## Output Formats
+
+**Always use `-F json` or `-F csv` when the output will be parsed by an agent.** The default format (`pretty`) is for human display and is difficult to parse programmatically.
+
+```bash
+# JSON — best for structured parsing
+bunx chcli -q "SELECT * FROM events LIMIT 5" -F json
+
+# CSV — good for tabular data
+bunx chcli -q "SELECT * FROM events LIMIT 5" -F csv
+
+# JSONL (one JSON object per line) — good for streaming/large results
+bunx chcli -q "SELECT * FROM events LIMIT 100" -F jsonl
+```
+
+Available format aliases: `json`, `jsonl`/`ndjson`, `jsoncompact`, `csv`, `tsv`, `pretty`, `vertical`, `markdown`, `sql`. Any native ClickHouse format name also works.
+
+See `references/formats.md` for the full format reference.
+
+## Common Workflows
+
+### Schema Discovery
+
+```bash
+# List all databases
+bunx chcli -q "SHOW DATABASES" -F json
+
+# List tables in current database
+bunx chcli -q "SHOW TABLES" -F json
+
+# List tables in a specific database
+bunx chcli -q "SHOW TABLES FROM analytics" -F json
+
+# Describe table schema
+bunx chcli -q "DESCRIBE TABLE events" -F json
+
+# Show CREATE TABLE statement
+bunx chcli -q "SHOW CREATE TABLE events"
+```
+
+### Data Exploration
+
+```bash
+# Row count
+bunx chcli -q "SELECT count() FROM events" -F json
+
+# Sample rows
+bunx chcli -q "SELECT * FROM events LIMIT 10" -F json
+
+# Column statistics
+bunx chcli -q "SELECT uniq(user_id), min(created_at), max(created_at) FROM events" -F json
+```
+
+### Data Extraction
+
+```bash
+# Extract to CSV file
+bunx chcli -q "SELECT * FROM events WHERE date = '2024-01-01'" -F csv > export.csv
+
+# Extract as JSON
+bunx chcli -q "SELECT * FROM events LIMIT 1000" -F json > export.json
+```
+
+## Additional Flags
+
+| Flag | Description |
+|------|-------------|
+| `-t, --time` | Print execution time to stderr |
+| `-v, --verbose` | Print query metadata (format, elapsed time) to stderr |
+| `--help` | Show help text |
+| `--version` | Print version |
+
+## Best Practices for Agents
+
+1. **Always specify `-F json` or `-F csv`** — never rely on the default format, which varies by TTY context.
+2. **Always use `LIMIT`** on SELECT queries unless you know the table is small. ClickHouse tables can contain billions of rows.
+3. **Start with schema discovery** — run `SHOW TABLES` and `DESCRIBE TABLE` before querying unfamiliar databases.
+4. **Use `-t` for timing** — helps gauge whether queries are efficient.
+5. **Prefer env vars for connection** — set them once in `.env` rather than repeating flags on every command.
+6. **Use `count()` first** — before extracting data, check how many rows match to avoid overwhelming output.
diff --git a/skills/clickhouse-query/references/connection.md b/skills/clickhouse-query/references/connection.md
new file mode 100644
index 0000000..3ec739c
--- /dev/null
+++ b/skills/clickhouse-query/references/connection.md
@@ -0,0 +1,94 @@
+# Connection Configuration Reference
+
+chcli connects to ClickHouse over HTTP(S). Connection details can be set via environment variables or CLI flags.
+
+## Precedence
+
+CLI flags take precedence over environment variables. If neither is set, the default value is used.
+
+```
+CLI flag > Environment variable > Default value
+```
+
+## Configuration Options
+
+| Flag | Env Var | Default | Description |
+|------|---------|---------|-------------|
+| `--host <host>` | `CLICKHOUSE_HOST` | `localhost` | ClickHouse server hostname or IP |
+| `--port <port>` | `CLICKHOUSE_PORT` | `8123` | HTTP interface port |
+| `-u, --user <user>` | `CLICKHOUSE_USER` | `default` | Authentication username |
+| `--password <pass>` | `CLICKHOUSE_PASSWORD` | *(empty)* | Authentication password |
+| `-d, --database <db>` | `CLICKHOUSE_DATABASE` | `default` | Default database for queries |
+| `-s, --secure` | `CLICKHOUSE_SECURE` | `false` | Use HTTPS instead of HTTP |
+
+## Connection URL
+
+chcli constructs the connection URL as:
+
+```
+{protocol}://{host}:{port}
+```
+
+Where `protocol` is `https` if `--secure` is set or `CLICKHOUSE_SECURE=true`, otherwise `http`.
+
+## Examples
+
+### Local Development (defaults)
+
+No configuration needed — connects to `http://localhost:8123` with user `default`:
+
+```bash
+bunx chcli -q "SELECT 1"
+```
+
+### Remote Instance via CLI Flags
+
+```bash
+bunx chcli \
+  --host ch.example.com \
+  --port 8443 \
+  --secure \
+  --user admin \
+  --password secret \
+  -d analytics \
+  -q "SELECT count() FROM events"
+```
+
+### Remote Instance via Environment Variables
+
+Create a `.env` file (Bun loads it automatically):
+
+```env
+CLICKHOUSE_HOST=ch.example.com
+CLICKHOUSE_PORT=8443
+CLICKHOUSE_SECURE=true
+CLICKHOUSE_USER=admin
+CLICKHOUSE_PASSWORD=secret
+CLICKHOUSE_DATABASE=analytics
+```
+
+Then run queries without connection flags:
+
+```bash
+bunx chcli -q "SELECT count() FROM events"
+```
+
+### ClickHouse Cloud
+
+ClickHouse Cloud uses HTTPS on port 8443:
+
+```env
+CLICKHOUSE_HOST=abc123.us-east-1.aws.clickhouse.cloud
+CLICKHOUSE_PORT=8443
+CLICKHOUSE_SECURE=true
+CLICKHOUSE_USER=default
+CLICKHOUSE_PASSWORD=your-password
+```
+
+### Mixed (Env Vars + Flag Override)
+
+Set base connection in `.env`, override database per-query:
+
+```bash
+bunx chcli -d other_db -q "SHOW TABLES"
+```
diff --git a/skills/clickhouse-query/references/formats.md b/skills/clickhouse-query/references/formats.md
new file mode 100644
index 0000000..f60278e
--- /dev/null
+++ b/skills/clickhouse-query/references/formats.md
@@ -0,0 +1,51 @@
+# Output Formats Reference
+
+chcli supports format aliases that map to ClickHouse native format names. Use the `-F` / `--format` flag to specify the output format.
+
+## Default Behavior
+
+- **TTY (interactive terminal):** `pretty` (PrettyCompactMonoBlock) — human-readable tables
+- **Piped/redirected output:** `tsv` (TabSeparatedWithNames) — machine-friendly tab-separated values
+
+## Format Alias Table
+
+| Alias | ClickHouse Format | Description |
+|-------|-------------------|-------------|
+| `json` | JSON | Full JSON object with `data` array, column metadata, row count, and statistics |
+| `jsonl` | JSONEachRow | One JSON object per line (newline-delimited JSON) |
+| `ndjson` | JSONEachRow | Alias for `jsonl` |
+| `jsoncompact` | JSONCompact | JSON with column names separate from row data (arrays instead of objects) |
+| `csv` | CSVWithNames | Comma-separated values with a header row |
+| `tsv` | TabSeparatedWithNames | Tab-separated values with a header row |
+| `pretty` | PrettyCompactMonoBlock | Human-readable bordered table |
+| `vertical` | Vertical | Each column on its own line (useful for wide rows) |
+| `markdown` | Markdown | Markdown-formatted table |
+| `sql` | SQLInsert | Output as SQL INSERT statements |
+
+Any format name not in this table is passed directly to ClickHouse, so you can use any native ClickHouse format (e.g., `Parquet`, `Arrow`, `Avro`).
+
+## Choosing a Format
+
+### For Agent/Programmatic Use
+
+- **`json`** — Best for structured parsing. Returns a complete JSON object:
+  ```json
+  {
+    "meta": [{"name": "count()", "type": "UInt64"}],
+    "data": [{"count()": "42"}],
+    "rows": 1,
+    "statistics": {"elapsed": 0.001, "rows_read": 100}
+  }
+  ```
+- **`jsonl`** — Best for large result sets or streaming. One JSON object per line:
+  ```
+  {"id": 1, "name": "Alice"}
+  {"id": 2, "name": "Bob"}
+  ```
+- **`csv`** — Good for tabular data and import/export workflows.
+
+### For Human Display
+
+- **`pretty`** — Default in terminals. Bordered table layout.
+- **`vertical`** — Useful when rows have many columns. Each row displayed vertically.
+- **`markdown`** — Useful for embedding query results in documentation.