Access MDB Reader Skill

Read .mdb / .accdb databases. Two modes, mirroring tool-pdf-reader and tool-chm-reader: stdout (info / schema / one table / a SELECT / a search — read as you go) and full-dump to a markdown file (token-cheap, re-readable via Read/Grep in this and future sessions).

Read-only by construction. Opens with ADO Mode=1 (adModeRead) and refuses any statement that isn't a SELECT. It cannot modify a database.

SKILL_DIR below = this skill's directory (skills/tool-mdb-reader/ in the repo, or ~/.claude/skills/tool-mdb-reader/).

Setup

Windows only. Needs pywin32 (pip install pywin32) and the Microsoft ACE OLEDB provider (the Access Database Engine — already present on machines running Access-backed engineering tools).

Bitness must match: 64-bit Python needs 64-bit ACE. pyodbc is not used — the daemon interpreter doesn't have it, and ACE is reachable via COM without any install.

Usage

1. Info — the first call

python "<SKILL_DIR>/mdb_tool.py" info "path/to/file.mdb"

Size, SHA-256, and every user table with row + column counts.

2. Schema

python "<SKILL_DIR>/mdb_tool.py" schema "file.mdb"                 # all tables
python "<SKILL_DIR>/mdb_tool.py" schema "file.mdb" --table CABLES  # one

Column names + ADO types (Double, VarWChar, SmallInt, …).

3. One table to stdout

python "<SKILL_DIR>/mdb_tool.py" table "file.mdb" SETTINGS

4. Ad-hoc SELECT

python "<SKILL_DIR>/mdb_tool.py" query "file.mdb" "SELECT CODTAB, IFVAL FROM [DEFTAB] WHERE CODTAB='INSU'"

Anything not starting with SELECT is refused. Bracket table names ([DEFTAB]) — Access needs it for names that collide with keywords.

5. Search every table

python "<SKILL_DIR>/mdb_tool.py" search "file.mdb" "jute" --limit 20

Case-insensitive across all columns of all tables; prints the matching rows with their 1-based row index.

6. Dump everything to markdown (token-saver)

python "<SKILL_DIR>/mdb_tool.py" dump "file.mdb" "path/to/output.md"

Every row of every table, a summary table with anchor links, per-table schema lines, and a SHA-256 of the source. Single-column text tables (report templates) render as fenced blocks rather than mangled markdown tables.

• Use when the database will be referenced more than once: dump once, then Read/Grep the .md.
• For a vault MDB, dump beside it as <name>-extracted.md (matches the *-extracted convention from the PDF/CHM skills).
• A 236 KB / 9-table / 493-row database dumps to ~24 KB of markdown in a second.

Two source-fidelity decisions (know these before trusting a dump)

Both are safe for reading and wrong for a write path:

1. Doubles render at 15 significant digits. The stored doubles carry IEEE-754 round-trip noise — a cell Access displays as 13.3 reads back through COM as 13.299999999999999. Rendering repr() makes the dump unreadable and breaks a grep for the value the user actually knows. .15g matches what Access and the owning application show. If you need bit-exact doubles, read the MDB, not the dump.
2. Trailing NUL record-padding is stripped. Some tools pad fixed-width records with NUL bytes. str.rstrip() will not remove them (NUL isn't whitespace), and a single such row makes the entire output file "binary" to grep — defeating the whole point. An interior NUL (never yet seen) becomes a visible ␀ rather than vanishing silently.

The dump's header states both, so a reader who didn't invoke the tool still knows.

Workflow for a legacy engineering database

1. info — how many tables, how big.
2. schema — column names + types; spot the lookup tables vs the data tables.
3. dump to a sibling *-extracted.md — the durable artifact.
4. search / query for specific values; Read/Grep the .md for everything else.

Notes

• Access exposes no foreign keys in many legacy files — relationships are usually inferred from the data, not declared. Don't assume the schema tells you the joins.
• Table names with spaces work ([My Table]); the dump's anchor links slugify them.
• This tool never writes. For a write path, the database's owning app is the right tool — Access rewrites index/catalog structures that a naive writer will corrupt.

Trigger phrases

User says	Action
"read this mdb" / "what's in this Access db"	info → schema → dump
"dump the mdb tables" / "save it as markdown"	dump to sibling *-extracted.md
"search the db for X"	search
"how many X in the database"	query with a SELECT