@shibbirweb/mcp-db-read-only 0.2.0 → 1.0.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/README.md CHANGED
@@ -3,80 +3,26 @@
3
3
  [![CI](https://github.com/shibbirweb/mcp-db-read-only/actions/workflows/ci.yml/badge.svg)](https://github.com/shibbirweb/mcp-db-read-only/actions/workflows/ci.yml)
4
4
  [![npm](https://img.shields.io/npm/v/%40shibbirweb%2Fmcp-db-read-only?label=npm&color=cb3837)](https://www.npmjs.com/package/@shibbirweb/mcp-db-read-only)
5
5
  [![Docker Hub](https://img.shields.io/docker/v/shibbirweb/mcp-db-read-only?label=docker%20hub&sort=semver)](https://hub.docker.com/r/shibbirweb/mcp-db-read-only)
6
- [![Image size](https://img.shields.io/docker/image-size/shibbirweb/mcp-db-read-only/latest?style=flat&label=image%20size)](https://hub.docker.com/r/shibbirweb/mcp-db-read-only/tags)
7
6
  [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](LICENSE)
8
7
 
9
- An [MCP](https://modelcontextprotocol.io) server that gives an AI assistant **read-only** access to your databases, whichever kind they are, and lets it **switch database, server, credentials and even engine mid-conversation without restarting the client**.
10
-
11
- | Engine | URL scheme | Queried with |
12
- | --- | --- | --- |
13
- | MySQL, MariaDB | `mysql://`, `mariadb://` | `run_query` (SQL) |
14
- | PostgreSQL (and wire-compatible) | `postgres://`, `postgresql://` | `run_query` (SQL) |
15
- | SQLite | `sqlite:///path/to/file.db` | `run_query` (SQL) |
16
- | SQL Server, Azure SQL | `mssql://`, `sqlserver://` | `run_query` (T-SQL) |
17
- | ClickHouse | `clickhouse://`, `clickhouse+https://` | `run_query` (SQL) |
18
- | MongoDB | `mongodb://`, `mongodb+srv://` | `find_documents`, `aggregate`, `count_documents`, `distinct_values` |
19
- | Redis (and Valkey, KeyDB) | `redis://`, `rediss://` | `redis_command` |
20
- | Elasticsearch, OpenSearch | `elasticsearch://`, `opensearch://`, `+https` variants | `search` |
21
-
22
- The browse tools (`list_tables`, `describe_table`, `get_table_sample`, ...) work on every engine, in that engine's terms: tables, collections, keys or indices.
23
-
24
- Runs from npm with `npx`, or entirely in Docker with nothing installed on your machine.
25
-
26
- ```mermaid
27
- flowchart LR
28
- A["AI assistant<br/>Claude Desktop / Claude Code"]
29
- B["mcp-db-read-only<br/>one process, whole session"]
30
- C[("PostgreSQL<br/>app")]
31
- D[("MySQL<br/>legacy")]
32
- E[("MongoDB<br/>events")]
33
- F[("Redis<br/>cache")]
34
- G[("anything reached<br/>with connect")]
35
-
36
- A <-->|"MCP over stdio"| B
37
- B -.->|"one driver per target"| C
38
- B -.->|"one driver per target"| D
39
- B -.->|"one driver per target"| E
40
- B -.->|"one driver per target"| F
41
- B -.->|"opened at runtime"| G
42
- ```
43
-
44
- The server lives for the whole session, so the active connection is just state inside it. Switching selects a different driver rather than reconnecting, and switching back reuses a warm one.
45
-
46
- ---
8
+ Let your AI assistant **look at your databases without being able to change them.**
47
9
 
48
- ## Quick start
49
-
50
- ### npm
10
+ Point it at MySQL, PostgreSQL, SQLite, SQL Server, ClickHouse, MongoDB, Redis or Elasticsearch, then just ask questions in plain words: *"how many users signed up this week?"*, *"what's in the orders table?"*, *"which Redis keys hold sessions?"*. It can read anything you give it access to, and it cannot write, update or delete anything.
51
11
 
52
- ```bash
53
- DB_URL='postgres://readonly:secret@127.0.0.1:5432/my_database' npx -y @shibbirweb/mcp-db-read-only
54
- ```
12
+ - **Every popular database, one server.** Switch between them in the middle of a conversation.
13
+ - **Read-only, twice over.** Every query is checked before it is sent, and the database itself is also told to refuse writes.
14
+ - **No restart to switch.** Change database, server or engine by asking.
15
+ - **Optional logging**, with a live page in your browser that shows every query as it happens.
55
16
 
56
- Requires Node 22.13 or newer. There is no container in the way, so `127.0.0.1` means what you expect.
17
+ Works with Claude Desktop, Claude Code, and any other [MCP](https://modelcontextprotocol.io) client.
57
18
 
58
- ### Docker
59
-
60
- ```bash
61
- docker run -i --rm \
62
- --add-host host.docker.internal:host-gateway \
63
- -e DB_URL='postgres://readonly:secret@host.docker.internal:5432/my_database' \
64
- shibbirweb/mcp-db-read-only
65
- ```
66
-
67
- Use `host.docker.internal` to reach a database on the same machine as Docker. Inside the container, `localhost` means the container itself.
68
-
69
- For SQLite in Docker, mount the file's directory read-only and point at the path inside the container:
70
-
71
- ```bash
72
- docker run -i --rm -v "$PWD/data:/data:ro" -e DB_URL='sqlite:///data/app.db' shibbirweb/mcp-db-read-only
73
- ```
19
+ ---
74
20
 
75
- The container is the more isolated of the two: the server runs with only what the image and the environment give it. Over npm it runs directly on your machine with your user's access. Both enforce the same read-only guarantees.
21
+ ## Quick start
76
22
 
77
- ### Claude Desktop
23
+ **1. Have Node.js 22.13 or newer** (`node --version`), or Docker.
78
24
 
79
- Add to `claude_desktop_config.json`:
25
+ **2. Add the server to your client.** For Claude Desktop, edit `claude_desktop_config.json` (Settings, Developer, Edit Config). For Claude Code, create `.mcp.json` in your project:
80
26
 
81
27
  ```json
82
28
  {
@@ -85,379 +31,308 @@ Add to `claude_desktop_config.json`:
85
31
  "command": "npx",
86
32
  "args": ["-y", "@shibbirweb/mcp-db-read-only"],
87
33
  "env": {
88
- "DB_PROFILES": "{\"app\": \"postgres://readonly@127.0.0.1/app\", \"cache\": \"redis://127.0.0.1:6379/0\"}",
89
- "DB_DEFAULT_PROFILE": "app"
34
+ "DB_URL": "postgres://readonly:secret@localhost:5432/myapp"
90
35
  }
91
36
  }
92
37
  }
93
38
  }
94
39
  ```
95
40
 
96
- Or the same server in Docker:
41
+ Replace the `DB_URL` with your own database. The examples below show one for every kind.
97
42
 
98
- ```json
99
- {
100
- "mcpServers": {
101
- "databases": {
102
- "command": "docker",
103
- "args": [
104
- "run", "-i", "--rm",
105
- "--add-host", "host.docker.internal:host-gateway",
106
- "-e", "DB_URL=postgres://readonly:secret@host.docker.internal:5432/app",
107
- "shibbirweb/mcp-db-read-only"
108
- ]
109
- }
110
- }
111
- }
112
- ```
43
+ **3. Restart the client once, and ask away:**
113
44
 
114
- ### Claude Code
45
+ > "What tables are in my database?"
46
+ > "Show me the 5 newest orders."
47
+ > "How many customers are in each country?"
115
48
 
116
- The same shape, in `.mcp.json` at your project root. Either form above works.
49
+ That's it. You never need to restart again to change database; just ask the assistant to switch.
117
50
 
118
- Restart the client once. After that you never need to restart it to change database.
51
+ ---
119
52
 
120
- > Credentials in these files sit on disk in plain text. Prefer read-only database accounts, and keep the file out of version control. See [Security](#security).
53
+ ## Examples for each database
121
54
 
122
- ### Coming from mcp-mysql-read-only
55
+ Each database has a URL **format**, then a real **example** to copy and change. Put the finished URL in `DB_URL`.
123
56
 
124
- This server reads the old `MYSQL_HOST`, `MYSQL_USER`, `MYSQL_PASSWORD`, `MYSQL_DATABASE`, `MYSQL_PROFILES` and `MYSQL_DEFAULT_PROFILE` unchanged, so swapping the image or package name is enough. Tool names are the same too; the one change is that `connect` now takes a URL.
57
+ Replace each `[PLACEHOLDER]` with your own value:
125
58
 
126
- ---
59
+ | Placeholder | What to put there |
60
+ | --- | --- |
61
+ | `[USER]` | The database user name |
62
+ | `[PASSWORD]` | That user's password |
63
+ | `[HOST]` | The server's address, e.g. `localhost` or `db.example.com` |
64
+ | `[PORT]` | The server's port. Optional: leave out `:[PORT]` to use the usual one shown for each database |
65
+ | `[DATABASE]` | The database name. Optional for most: leave it out and ask the assistant to list them |
127
66
 
128
- ## Switching connections
67
+ No password? Leave out `:[PASSWORD]`. No user either? Leave out `[USER]:[PASSWORD]@` entirely.
129
68
 
130
- Just ask. These map onto the connection tools:
69
+ ### MySQL and MariaDB
131
70
 
132
- > "switch to the staging database"
133
- > "what collections are in the events database?"
134
- > "connect to the Redis on 10.0.0.5 and show me the session keys"
71
+ Format (usual port 3306):
135
72
 
136
- | Want | Restart? |
137
- | --- | --- |
138
- | Another database on the same server | No |
139
- | Another named profile, on any engine | No |
140
- | A different server, credentials or engine | No |
141
- | A new permanent profile in `DB_PROFILES` | Yes, once |
142
-
143
- ```mermaid
144
- sequenceDiagram
145
- autonumber
146
- actor You
147
- participant A as Assistant
148
- participant S as MCP server
149
- participant P as PostgreSQL
150
- participant M as MongoDB
151
-
152
- You->>A: "how many signups yesterday?"
153
- A->>S: run_query(SELECT count(*) ...)
154
- S->>P: read-only transaction
155
- P-->>S: 4821
156
- A-->>You: 4821 signups
157
-
158
- You->>A: "and how many of them opened the app?"
159
- A->>S: use_connection(events)
160
- S->>M: connect + ping
161
- Note over S: verified, so the switch is committed
162
- A->>S: count_documents(opens, {...})
163
- S->>M: countDocuments
164
- M-->>S: 3907
165
- A-->>You: 3907 of them
73
+ ```text
74
+ mysql://[USER]:[PASSWORD]@[HOST]:[PORT]/[DATABASE]
75
+ mariadb://[USER]:[PASSWORD]@[HOST]:[PORT]/[DATABASE]
166
76
  ```
167
77
 
168
- A switch that fails verification is never committed, so the previous connection stays active and the session keeps working.
78
+ Example:
169
79
 
170
- ### Named profiles
80
+ ```text
81
+ mysql://readonly:secret@localhost:3306/shop
82
+ ```
171
83
 
172
- Define several connections up front with `DB_PROFILES`, a JSON object whose values are URLs, or objects with a separate password:
84
+ > "List the tables in shop." · "Describe the orders table." · "What were last month's top 10 products by revenue?"
173
85
 
174
- ```json
175
- {
176
- "app": "postgres://readonly@db.internal:5432/app",
177
- "legacy": "mysql://readonly@legacy.internal/shop",
178
- "events": { "url": "mongodb://reader@mongo.internal/events", "password": "p@ss/w#rd" },
179
- "cache": "redis://cache.internal:6379/0",
180
- "logs": "elasticsearch+https://reader@logs.internal:9200",
181
- "reports": "sqlite:///data/reports.db"
182
- }
86
+ ### PostgreSQL
87
+
88
+ Format (usual port 5432). Add `?sslmode=require` to use TLS:
89
+
90
+ ```text
91
+ postgres://[USER]:[PASSWORD]@[HOST]:[PORT]/[DATABASE]
183
92
  ```
184
93
 
185
- The object form exists because a password inside a URL must be percent-encoded, and one containing `@`, `/` or `#` otherwise splits the URL in the wrong place. A profile that fails to parse is skipped with a warning rather than taking the server down.
94
+ Example:
186
95
 
187
- ### Reaching somewhere not in the profiles
96
+ ```text
97
+ postgres://readonly:secret@localhost:5432/myapp
98
+ ```
188
99
 
189
- The `connect` tool takes a URL (and optionally a separate password) at runtime and keeps it for the rest of the session under an alias. Nothing is written to disk, and no restart is involved.
100
+ > "Which tables are in the reporting schema?" · "Show the foreign keys on invoices." · "Count signups per day this week."
190
101
 
191
- ### URL details
102
+ ### SQLite
192
103
 
193
- | Engine | Notes |
194
- | --- | --- |
195
- | MySQL | `?ssl=true` requires TLS; `?ssl-mode=VERIFY_IDENTITY` also checks the certificate |
196
- | PostgreSQL | `?sslmode=require`, `verify-ca` or `verify-full`, as in libpq |
197
- | SQLite | `sqlite:///absolute/path.db`; a relative path is resolved once, at startup |
198
- | SQL Server | Encrypted by default. `?trustServerCertificate=true` for self-signed development servers; `?applicationIntent=ReadOnly` routes to a readable secondary |
199
- | ClickHouse | The HTTP interface: port 8123, or 8443 with `clickhouse+https` |
200
- | MongoDB | Replica sets as `mongodb://a:27017,b:27017/db?replicaSet=rs0`; any driver option passes through the query string |
201
- | Redis | The path is the database number: `redis://host:6379/3` |
202
- | Elasticsearch | `?api_key=...` authenticates with an API key; it is treated as a secret and never displayed |
104
+ Format (three slashes, then the full path to the file):
203
105
 
204
- ---
106
+ ```text
107
+ sqlite:///[PATH_TO_FILE]
108
+ ```
205
109
 
206
- ## Tools
110
+ Example:
207
111
 
208
- ### Connection
112
+ ```text
113
+ sqlite:///Users/me/data/app.db
114
+ ```
209
115
 
210
- | Tool | Purpose |
211
- | --- | --- |
212
- | `current_connection` | Which engine, server and database is active |
213
- | `list_connections` | Available profiles and their engines, `*` marks the active one |
214
- | `list_databases` | Databases on the connected server |
215
- | `use_database` | Switch database on the current server |
216
- | `use_connection` | Switch to a named profile, optional `database` override |
217
- | `connect` | Open any server from a URL, optional `alias` |
116
+ The file is opened read-only.
218
117
 
219
- ### Browsing, on every engine
118
+ > "What tables does this file have?" · "Show 10 rows from notes."
220
119
 
221
- | Tool | SQL engines | MongoDB | Redis | Elasticsearch |
222
- | --- | --- | --- | --- | --- |
223
- | `list_tables` | tables and views | collections | keys, by SCAN | indices |
224
- | `describe_table` | columns | fields inferred from 100 sampled documents | type, TTL, length | mapping |
225
- | `get_table_indexes` | indexes | indexes | n/a | n/a |
226
- | `get_foreign_keys` | foreign keys (not ClickHouse) | n/a | n/a | n/a |
227
- | `get_table_sample` | up to 50 rows | up to 50 documents | the start of the value | up to 50 hits |
120
+ ### SQL Server (and Azure SQL)
228
121
 
229
- `list_tables` takes an optional glob `pattern`, such as `user*`, which is how you browse a Redis instance with millions of keys.
122
+ Format (usual port 1433). Add `?trustServerCertificate=true` for a local server with a self-signed certificate:
230
123
 
231
- ### Querying, per engine family
124
+ ```text
125
+ mssql://[USER]:[PASSWORD]@[HOST]:[PORT]/[DATABASE]
126
+ ```
232
127
 
233
- | Tool | Engine | Accepts |
234
- | --- | --- | --- |
235
- | `run_query` | SQL engines | One read-only statement in the engine's own dialect |
236
- | `find_documents` | MongoDB | Filter, projection, sort, limit and skip, as Extended JSON |
237
- | `aggregate` | MongoDB | A pipeline, without `$out` or `$merge` |
238
- | `count_documents` | MongoDB | A filter |
239
- | `distinct_values` | MongoDB | A field and an optional filter |
240
- | `search` | Elasticsearch, OpenSearch | A Query DSL body; `size: 0` with `track_total_hits` counts |
241
- | `redis_command` | Redis | One read-only command and its arguments |
128
+ Example:
242
129
 
243
- Every tool is advertised all the time, since MCP fixes the tool list at startup while the active engine can change. Calling one against the wrong engine says which tools fit instead.
130
+ ```text
131
+ mssql://readonly:secret@localhost:1433/Sales?trustServerCertificate=true
132
+ ```
244
133
 
245
- Every reading tool also accepts an optional `database`, applied to that call only, leaving the active connection alone.
134
+ > "List the tables in Sales." · "Show the top 5 customers by order total." (SQL Server uses `TOP 5`, not `LIMIT`; the assistant knows.)
246
135
 
247
- ---
136
+ ### ClickHouse
248
137
 
249
- ## Configuration
138
+ Format (usual port 8123, or 8443 with `clickhouse+https`):
250
139
 
251
- | Variable | Default | Purpose |
252
- | --- | --- | --- |
253
- | `DB_URL` | none | One connection, registered as the profile `default` |
254
- | `DB_PASSWORD` | none | Password for `DB_URL`, so it need not be encoded into the URL |
255
- | `DB_PROFILES` | none | JSON object of named profiles |
256
- | `DB_DEFAULT_PROFILE` | none | Which profile starts active |
257
- | `DB_QUERY_TIMEOUT_MS` | `30000` | Statement timeout, enforced by each server where it can be |
258
- | `DB_CONNECT_TIMEOUT_MS` | `10000` | Connection timeout |
259
- | `DB_LOG` | off | `true` logs every tool call to stderr. See [Call logging](#call-logging) |
260
- | `DB_LOG_FILE` | none | Log every tool call to this file instead |
261
- | `DB_LOG_DIR` | none | Save every tool call as its own JSON file in this folder, permanently |
262
- | `DB_LOG_FORMAT` | `pretty` | `pretty` or `json` |
263
- | `DB_LOG_PORT` | none | Serve a live log viewer in the browser on this port |
264
- | `DB_LOG_HISTORY` | `500` | Entries the viewer keeps in memory when there is no `DB_LOG_DIR` |
265
- | `MYSQL_*` | | The legacy MySQL-only variables, read unchanged. See above |
140
+ ```text
141
+ clickhouse://[USER]:[PASSWORD]@[HOST]:[PORT]/[DATABASE]
142
+ clickhouse+https://[USER]:[PASSWORD]@[HOST]:[PORT]/[DATABASE]
143
+ ```
266
144
 
267
- None of these are required: with no configuration at all the server still starts, and the tools tell you to call `connect`.
145
+ Example:
268
146
 
269
- Starting profile: `DB_DEFAULT_PROFILE` (or `MYSQL_DEFAULT_PROFILE`) if it names a real profile, else `default`, else the first one defined.
147
+ ```text
148
+ clickhouse://reader:secret@localhost:8123/analytics
149
+ ```
270
150
 
271
- ---
151
+ > "How many events per hour did we have yesterday?" · "What is the sorting key of the events table?"
272
152
 
273
- ## Call logging
153
+ ### MongoDB
274
154
 
275
- Off by default. Turn it on to see every tool call the assistant makes: its input, each statement the drivers sent to the database, and the full output.
155
+ Format (usual port 27017). Use `mongodb+srv` for MongoDB Atlas, with no port:
276
156
 
277
- | Variable | Effect |
278
- | --- | --- |
279
- | `DB_LOG=true` | Log to stderr, which your MCP client keeps (Claude Code shows it in its MCP logs) |
280
- | `DB_LOG_FILE=/path/calls.log` | Log to that file instead, appended to, created readable by you only. Implies `DB_LOG` |
281
- | `DB_LOG_DIR=/path/folder` | Save every entry as its own JSON file in that folder, permanently. Implies logging on its own, without also writing text |
282
- | `DB_LOG_FORMAT=json` | One JSON object per line, for `jq` or a log shipper. The default is `pretty` |
157
+ ```text
158
+ mongodb://[USER]:[PASSWORD]@[HOST]:[PORT]/[DATABASE]?authSource=admin
159
+ mongodb+srv://[USER]:[PASSWORD]@[CLUSTER_HOST]/[DATABASE]
160
+ ```
283
161
 
284
- A pretty entry:
162
+ Example:
285
163
 
286
164
  ```text
287
- ┌─ #3 run_query · ok · 38 ms · 2026-09-25T10:14:03.221Z
288
- │ connection dev (MySQL) mysql://reader@127.0.0.1:3306/app
289
- │ input
290
- │ {
291
- │ "query": "SELECT COUNT(*) AS n FROM members"
292
- │ }
293
- │ statements (1)
294
- │ 1. MySQL · 12 ms · 1 row
295
- │ SELECT COUNT(*) AS n FROM members
296
- │ output
297
- │ [
298
- │ {
299
- │ "n": 17440
300
- │ }
301
- │ ]
302
- └─
165
+ mongodb://reader:secret@localhost:27017/myapp?authSource=admin
303
166
  ```
304
167
 
305
- - **Everything is logged, including the full output of every call.** Treat a log file like the data it contains.
306
- - **Credentials never are.** A `password` argument, the password inside a connection URL, and secret-looking URL options such as `api_key` are always written as `***`, with no way to turn that off.
307
- - **Statements** include the ones the server sends on its own behalf: PostgreSQL's `BEGIN READ ONLY` and `ROLLBACK`, Redis's `COMMAND INFO` checks, the catalog queries behind `describe_table`. Each shows its duration and outcome (a row count, or the error); the data itself is in the call's output. A statement sent outside any call, such as MySQL's per-connection setup, gets an entry of its own.
308
- - An entry is written when its call finishes, and numbered when it starts, so calls handled concurrently can appear out of numeric order.
309
- - With `DB_LOG_DIR`, each entry is its own file in a folder per UTC day, e.g. `2026-09-25/103014-221Z_p72440_c000012_run_query_ok.json`, holding the full record as pretty JSON: time, pid, sequence, tool and outcome are in the name, so `ls` and `grep` work without opening anything. Files and folders are readable by you only. Nothing is ever deleted or rotated; archive or remove old day folders yourself. Every copy of the server can share one folder, since names never collide.
310
- - Logging can never break a call. If the log cannot be written (say, the disk is full), the call still succeeds, one warning is printed, and logging stops.
168
+ > "What collections are in myapp?" · "What fields do documents in users have?" · "Find the 5 most recent orders over 100." · "Count users by country."
311
169
 
312
- In Docker, a log file or folder must be on a mounted volume to outlive the container: `-v "$PWD/logs:/logs" -e DB_LOG_DIR=/logs`. `DB_LOG=true` needs no mount.
170
+ ### Redis (and Valkey, KeyDB)
313
171
 
314
- ### Live viewer in the browser
172
+ Format (usual port 6379). `[DB_NUMBER]` is the database number, 0 if left out; `rediss` means TLS:
315
173
 
316
- Add `DB_LOG_PORT` to watch calls arrive in a browser page, updating the moment each one finishes:
174
+ ```text
175
+ redis://[USER]:[PASSWORD]@[HOST]:[PORT]/[DB_NUMBER]
176
+ rediss://[USER]:[PASSWORD]@[HOST]:[PORT]/[DB_NUMBER]
177
+ ```
317
178
 
318
- ```bash
319
- DB_LOG=true DB_LOG_PORT=4800 npx -y @shibbirweb/mcp-db-read-only
320
- # then open http://127.0.0.1:4800/
179
+ Example:
180
+
181
+ ```text
182
+ redis://localhost:6379/0
321
183
  ```
322
184
 
323
- | Variable | Default | Effect |
324
- | --- | --- | --- |
325
- | `DB_LOG_PORT` | none | Serve the live viewer on this port. Needs `DB_LOG_DIR`, `DB_LOG` or `DB_LOG_FILE` as well; alone it only prints a warning |
326
- | `DB_LOG_HISTORY` | `500` | Without `DB_LOG_DIR`, how many recent entries the viewer keeps in memory. `0` keeps none |
185
+ > "Which keys start with session:?" · "What's inside user:42?" · "How long until cache:home expires?"
327
186
 
328
- The page lists calls newest first, **20 per page by default**, with page controls and a choice of 10, 20, 30 or 50 per page (remembered in your browser). Each call is a card (tool, status, duration, connection, and which client and process made it) that expands to its input, statements and output, every block with its own copy icon. Filters (text, tool, failures only) apply across everything logged, not just the page shown. Page 1 updates itself as calls arrive; on any other page a "new entries" button appears instead, so the page you are reading does not shift. It follows your system's light or dark theme and loads nothing from the internet.
187
+ ### Elasticsearch and OpenSearch
329
188
 
330
- With `DB_LOG_DIR` the viewer reads the folder, so it shows everything ever saved there, across restarts, **including calls made by other copies of the server** sharing the folder (Claude Desktop runs several). Without it, it shows this copy's last `DB_LOG_HISTORY` calls from memory.
189
+ Format (usual port 9200, no database). Add `+https` for TLS, or `?api_key=[API_KEY]` instead of a user and password:
331
190
 
332
- > **The viewer has no access control and listens on every network interface (`0.0.0.0`).** Anyone who can reach the port, including other devices on the same network, can read every query and every result in the log. Use it on a network you trust, or block the port at your firewall. The server prints a reminder of this when the viewer starts.
191
+ ```text
192
+ elasticsearch://[USER]:[PASSWORD]@[HOST]:[PORT]
193
+ opensearch://[USER]:[PASSWORD]@[HOST]:[PORT]
194
+ ```
333
195
 
334
- With no `DB_LOG_PORT`, nothing listens on any port.
196
+ Example:
335
197
 
336
- The viewer takes its port on the **first tool call**, not at startup. MCP clients such as Claude Desktop start one copy of the server per chat surface, and most copies are never used; binding lazily means the copy your chat is using gets the port, and idle copies hold none (they open no database connections either until used).
198
+ ```text
199
+ elasticsearch+https://elastic:secret@search.example.com:9200
200
+ ```
337
201
 
338
- If the port is taken when a call arrives (another chat is already using the viewer, say), the call still works, and its result carries one extra line saying which process holds the port: free it, or set `DB_LOG_PORT` to another one. Every later call retries, so once the port is free the viewer comes up on the next call and says so. `current_connection` always shows the viewer's state.
202
+ > "What indices do we have?" · "Find error logs from the last hour." · "How many documents are in logs-2026.09?"
339
203
 
340
- In Docker, publish the port as well: `-p 4800:4800 -e DB_LOG=true -e DB_LOG_PORT=4800`. Publishing it as `-p 127.0.0.1:4800:4800` keeps it reachable from this machine only.
204
+ Detailed notes for every database, including how to create a read-only account, are in the [Databases guide](https://github.com/shibbirweb/mcp-db-read-only/wiki/Databases).
341
205
 
342
206
  ---
343
207
 
344
- ## Security
208
+ ## Using several databases
345
209
 
346
- Every engine is kept read-only by **two independent layers**, so a hole in one is not automatically a write. The first layer runs before any connection is used; the second is enforced by the database server itself wherever the engine offers a way, and structurally where it does not.
210
+ Give each one a name with `DB_PROFILES`, and switch by asking ("switch to legacy", "use the cache"):
347
211
 
348
- | Engine | Layer one, in this server | Layer two |
349
- | --- | --- | --- |
350
- | MySQL, MariaDB | SQL validator | `SET SESSION TRANSACTION READ ONLY` on every connection; the driver cannot send a second statement |
351
- | PostgreSQL | SQL validator, dialect-aware (dollar quotes, `E''` strings) | Every statement runs in a `READ ONLY` transaction that is always rolled back; extended query protocol, one statement only |
352
- | SQLite | SQL validator | The file is opened read-only by SQLite; extensions disabled |
353
- | SQL Server | SQL validator, scanning every statement since T-SQL needs no separators | Every batch runs in a transaction that is always rolled back |
354
- | ClickHouse | SQL validator, refusing table functions that reach outside the server | ClickHouse's own `readonly` setting on every query |
355
- | MongoDB | Operator denylist: `$out`, `$merge`, `$function`, `$where` anywhere | Stage allowlist in the driver, which only ever calls read operations |
356
- | Redis | Command allowlist | The server's own `COMMAND INFO` flags: a command is sent only if Redis itself calls it read-only |
357
- | Elasticsearch | Search body allowlist; index names cannot address an API | The driver can only reach fixed read endpoints |
212
+ ```json
213
+ "env": {
214
+ "DB_PROFILES": "{\"app\": \"postgres://readonly@localhost/app\", \"legacy\": \"mysql://readonly@localhost/shop\", \"cache\": \"redis://localhost:6379/0\"}",
215
+ "DB_DEFAULT_PROFILE": "app"
216
+ }
217
+ ```
358
218
 
359
- The SQL validator lexes each dialect exactly as the server will: string literals, quoted identifiers and comments are blanked before any rule looks at the statement, so a keyword or semicolon inside a literal is never mistaken for SQL. Constructs it cannot be certain the server reads the same way, such as nested block comments or MySQL's executable `/*! */` comments, are refused rather than guessed at. Only the dialect's read statements may lead (`SELECT`, `WITH`, and `SHOW`, `DESCRIBE` or `EXPLAIN` where they exist). Functions that write files, reach other servers or run SQL hidden in a string (`INTO OUTFILE`, `lo_export`, `dblink`, `OPENROWSET`, ClickHouse's `url()` and `file()`) are blocked.
219
+ You can also connect to a database you didn't list, mid-conversation: *"connect to postgres://readonly@10.0.0.5/reports"*.
360
220
 
361
- Integration tests prove layer two separately: they send writes straight to each driver, bypassing every validator, and assert the server refused or undid them.
221
+ **Password with special characters** (`@`, `/`, `#`)? Give it separately instead of inside the URL: `DB_PASSWORD` next to `DB_URL`, or `{"url": "...", "password": "..."}` inside `DB_PROFILES`.
362
222
 
363
- ### What this is not
223
+ ---
364
224
 
365
- **This is a guard, not a permission system.** It stops an assistant from writing through *this* server. It does not stop anyone holding the same credentials from writing through any other client.
225
+ ## Running with Docker
366
226
 
367
- **Point it at read-only accounts.** This is the real protection, and on MongoDB and Elasticsearch, whose servers have no read-only session mode, it is the only server-side one:
227
+ Nothing to install but Docker:
368
228
 
369
- ```sql
370
- -- MySQL
371
- CREATE USER 'readonly'@'%' IDENTIFIED BY '...'; GRANT SELECT ON app.* TO 'readonly'@'%';
372
- -- PostgreSQL
373
- CREATE ROLE readonly LOGIN PASSWORD '...'; GRANT pg_read_all_data TO readonly;
229
+ ```json
230
+ {
231
+ "mcpServers": {
232
+ "databases": {
233
+ "command": "docker",
234
+ "args": [
235
+ "run", "-i", "--rm",
236
+ "--add-host", "host.docker.internal:host-gateway",
237
+ "-e", "DB_URL=postgres://readonly:secret@host.docker.internal:5432/myapp",
238
+ "shibbirweb/mcp-db-read-only"
239
+ ]
240
+ }
241
+ }
242
+ }
374
243
  ```
375
244
 
376
- ```js
377
- // MongoDB
378
- db.createUser({ user: "reader", pwd: "...", roles: [{ role: "read", db: "app" }] });
379
- ```
245
+ Inside Docker, use `host.docker.internal` instead of `localhost` to reach a database on your own computer. For SQLite, mount the folder: add `"-v", "/Users/me/data:/data:ro"` and use `sqlite:///data/app.db`.
380
246
 
381
- ```text
382
- # Redis
383
- ACL SETUSER reader on >... ~* +@read -@dangerous
247
+ ---
248
+
249
+ ## Watching what the assistant does
250
+
251
+ Turn on logging to keep a record of every query the assistant runs, and see them live in your browser.
252
+
253
+ **1. Save logs to a folder** by adding this to the server's `env`:
254
+
255
+ ```json
256
+ "DB_LOG_DIR": "/Users/me/Library/Logs/mcp-db-read-only"
384
257
  ```
385
258
 
386
- Other limits worth knowing:
259
+ Each query is saved as its own file, one folder per day. Nothing is ever deleted automatically.
387
260
 
388
- - Results are truncated to 100 rows in the tool output. Add a `LIMIT` (or `$limit`) when reading large tables.
389
- - A column named exactly like a write keyword must be quoted where the validator scans for them: inside `WITH` queries, and in every SQL Server statement.
390
- - SQLite queries run in a separate process, so one that exceeds the timeout can be killed outright.
261
+ **2. Open the viewer** in a terminal, whenever you want to watch:
391
262
 
392
- ---
263
+ ```bash
264
+ npx -y @shibbirweb/mcp-db-read-only viewer --dir /Users/me/Library/Logs/mcp-db-read-only --port 4800
265
+ ```
393
266
 
394
- ## Known behaviour
267
+ Then open **http://127.0.0.1:4800/**. You'll see every call: what was asked, the exact query sent, how long it took, and the result. It updates live, shows 20 per page (10, 20, 30 or 50 to choose from), and lets you filter and copy anything. Press Ctrl+C to close it.
395
268
 
396
- **Parallel tool calls.** The active connection is a single piece of process state. If a client issues several tool calls in one batch they are handled concurrently, so a `use_database` batched alongside a query is not guaranteed to land first. When a read must be pinned to a particular database, pass the per-call `database` argument instead.
269
+ > The viewer has no password. Anyone who can reach that port on your network can read the log while it runs.
397
270
 
398
- **Shutdown.** The server exits on `SIGINT`/`SIGTERM`, not when stdin closes. The live viewer, if running, closes its port with it. Open sockets keep the event loop alive, and stdin reaching EOF only means no further requests were buffered.
271
+ Passwords are never written to the logs. More in the [Logging guide](https://github.com/shibbirweb/mcp-db-read-only/wiki/Logging-and-Viewer).
399
272
 
400
273
  ---
401
274
 
402
- ## Development
275
+ ## Is it really read-only?
403
276
 
404
- Everything runs in Docker, so a clone and Docker are the only requirements:
277
+ Yes, in two independent ways, so a mistake in one is caught by the other:
405
278
 
406
- ```bash
407
- git clone https://github.com/shibbirweb/mcp-db-read-only.git
408
- cd mcp-db-read-only
409
- ./scripts/test-in-docker.sh # every engine
410
- ENGINES="postgres redis" ./scripts/test-in-docker.sh # a subset
411
- ```
279
+ 1. **Before anything is sent**, every query is checked. Only reads are allowed: `SELECT` and friends for SQL, read commands for Redis, searches for Elasticsearch, and no `$out` or `$merge` for MongoDB.
280
+ 2. **The database is told to refuse writes too**, wherever it supports that: read-only sessions on MySQL, read-only transactions on PostgreSQL, a read-only file on SQLite, and so on.
412
281
 
413
- That starts a throwaway container per engine, builds the test image, runs the full suite against them and tears everything down. Your own databases are never touched.
282
+ **The best protection is still a read-only database account.** Then nothing can write through it, whatever happens. The [Databases guide](https://github.com/shibbirweb/mcp-db-read-only/wiki/Databases) shows how to create one for each database.
414
283
 
415
- With Node 22.13 or newer installed locally:
284
+ Keep in mind that the assistant **can read** whatever the account can see, and what it reads becomes part of your conversation with the AI provider. Only connect accounts that can see data you are happy to share. See [PRIVACY.md](PRIVACY.md).
416
285
 
417
- ```bash
418
- npm ci
419
- npm run build
420
- npm run test:unit # no database needed
421
- npm test # integration suites skip any engine they cannot reach
422
- ```
423
-
424
- The SQLite and handshake suites need no server and always run. The others read `TEST_MYSQL_URL`, `TEST_POSTGRES_URL`, `TEST_MSSQL_URL`, `TEST_CLICKHOUSE_URL`, `TEST_MONGODB_URL`, `TEST_REDIS_URL` and `TEST_ELASTICSEARCH_URL`, and create and drop scratch data named `mcp_test*`, so point them at disposable servers.
286
+ ---
425
287
 
426
- ### Project structure
288
+ ## Settings
427
289
 
428
- ```
429
- src/
430
- index.ts Entry point
431
- ApplicationFactory.ts Composition root: the only file that wires things, and the only one naming a driver
432
- types/ Interfaces and type aliases, one file per concern
433
- errors/ Named error classes
434
- domain/ Engine catalog, ConnectionTarget, ConnectionProfile
435
- config/ Reading configuration from the environment
436
- connections/ URL parser, target factory, profile registry, connection manager
437
- drivers/ DatabaseDriver strategy, registry, LRU cache, and sql/ document/ keyvalue/ search/
438
- validation/ sql/ (dialects, skeletonizer, rules), document/, keyvalue/, search/, names/
439
- formatting/ Response, row and JSON rendering
440
- tools/ BaseTool, DatabaseScopedTool, connection/ browse/ sql/ document/ search/ keyvalue/
441
- server/ McpDbServer
442
- ```
290
+ All optional. Set them in the server's `env`.
443
291
 
444
- Developer documentation, including why each part is built the way it is, lives in the [wiki](https://github.com/shibbirweb/mcp-db-read-only/wiki) (source in [`docs/wiki/`](docs/wiki/)).
292
+ | Setting | What it does |
293
+ | --- | --- |
294
+ | `DB_URL` | The database to connect to at startup |
295
+ | `DB_PASSWORD` | The password for `DB_URL`, if you'd rather keep it out of the URL |
296
+ | `DB_PROFILES` | Several named databases, as JSON, to switch between |
297
+ | `DB_DEFAULT_PROFILE` | Which of those to start with |
298
+ | `DB_QUERY_TIMEOUT_MS` | Stop a query after this long (default 30000, that is 30 seconds) |
299
+ | `DB_CONNECT_TIMEOUT_MS` | Give up connecting after this long (default 10000) |
300
+ | `DB_LOG_DIR` | Save every call as a file in this folder |
301
+ | `DB_LOG=true` | Write every call to the client's log instead |
302
+ | `DB_LOG_FILE` | Write every call to one file instead |
303
+ | `DB_LOG_FORMAT` | `pretty` (default) or `json`, for `DB_LOG` and `DB_LOG_FILE` |
304
+ | `DB_LOG_PORT` | Run the viewer inside the server itself (the separate `viewer` command is usually better) |
305
+
306
+ Coming from `mcp-mysql-read-only`? Its `MYSQL_HOST`, `MYSQL_USER`, `MYSQL_PASSWORD`, `MYSQL_DATABASE` and `MYSQL_PROFILES` settings still work as they are.
307
+
308
+ Full details: [Configuration guide](https://github.com/shibbirweb/mcp-db-read-only/wiki/Configuration).
445
309
 
446
310
  ---
447
311
 
448
- ## Contributing
312
+ ## Troubleshooting
313
+
314
+ **"Can't connect" from Docker to a database on my computer.** Use `host.docker.internal` instead of `localhost`, and keep the `--add-host` line from the Docker example.
315
+
316
+ **The server won't start with npx.** Check `node --version` is 22.13 or newer.
449
317
 
450
- Pull requests target `master`. CI runs the full suite against every engine, twice (once with MySQL, once with MariaDB), and builds the image for amd64 and arm64. Please keep changes covered by tests, and update `docs/wiki/` when behaviour changes.
318
+ **SQL Server says the certificate isn't trusted.** Add `?trustServerCertificate=true` to the URL for a local or development server.
451
319
 
452
- There is a second copy of this document, [`README.dockerhub.md`](README.dockerhub.md), which is published as the Docker Hub description. Docker Hub renders neither mermaid nor relative links, so that copy uses ASCII diagrams and absolute URLs. **If you change user-facing behaviour here, change it there too.**
320
+ **"No database selected".** Your URL has no database name. Add one (`.../myapp`), or ask the assistant to list the databases and pick one.
453
321
 
454
- ## Changelog
322
+ **My password has `@` or `#` in it.** Use `DB_PASSWORD`, or the `{"url": ..., "password": ...}` form in `DB_PROFILES`.
455
323
 
456
- Release history is in [CHANGELOG.md](CHANGELOG.md).
324
+ **The viewer says the port is in use.** Something else is using it. Close that, or pick another port with `--port 4801`.
325
+
326
+ More answers in the [Troubleshooting guide](https://github.com/shibbirweb/mcp-db-read-only/wiki/Troubleshooting).
327
+
328
+ ---
457
329
 
458
- ## Privacy
330
+ ## Learn more
459
331
 
460
- The server sends nothing anywhere except to the databases you point it at: no telemetry, no analytics, nothing written to disk, nothing kept after it exits. What does leave your machine is whatever your assistant reads, since query results become conversation content. [PRIVACY.md](PRIVACY.md) sets out both halves.
332
+ - [Wiki](https://github.com/shibbirweb/mcp-db-read-only/wiki): user guides for every feature, plus developer documentation
333
+ - [CHANGELOG.md](CHANGELOG.md): what changed in each version
334
+ - [PRIVACY.md](PRIVACY.md): what the server sends where (nothing, except to your databases)
335
+ - Contributing: pull requests to `master` are welcome. `./scripts/test-in-docker.sh` runs the full test suite against every database in throwaway containers; see [Testing](https://github.com/shibbirweb/mcp-db-read-only/wiki/Testing). If you change this README, change [README.dockerhub.md](README.dockerhub.md) to match.
461
336
 
462
337
  ## License
463
338