Ask your AI agent about any binary open in IDA Pro, in plain English. It answers from the live database.
Tip
You don't need to know SQL. You ask in plain English: "who calls this function?", "rename these and comment them", "what does this binary do?". Your agent writes and runs the SQL for you. SQL is simply the language idasql speaks with the agent, one every agent already knows. You can type it yourself whenever you want to, but you never have to.
Note
idasql exports nothing. Despite the "sql" in its name, idasql never converts or exports your IDB to a SQL database, or to anything else. It is a virtual SQL interface on top of the IDA SDK: every query reads your live IDB, and every change goes straight back into it.
IDASQL gives AI coding agents full, live access to IDA databases: 70+ tables covering functions, xrefs, strings, types, pseudocode and more, all read and write. It works with Claude Code, Copilot CLI, Codex, or any MCP (Model Context Protocol) client, or anything that speaks HTTP. It is created by Elias Bachaalany.
▶ Watch: Vibe Reverse Engineering #1: IDASQL & the xSQL Concept
Work fully headless, with no IDA window at all, or right inside the IDA you already have open. The agent handles both.
/idasql:connect Open sample.exe in the background and let's analyze it together.
The agent runs IDA in the background (idasql -s) and analyses raw binaries itself. It opens as
many databases as you like, and saves them when you say so.
There's nothing to set up in IDA. Open your database as usual, then ask your agent. It finds your IDA by itself:
You: how many functions in target.dll?
The agent then runs:
$ idasql instances # every running IDA: target.dll.i64 is pid 4242, no SQL server
$ idasql http start 4242
remote control is off for target.dll.i64 (pid 4242)
turn it on with: idasql remote on 4242 (or .remote on in that IDA)
$ idasql remote on 4242 && idasql http start 4242
remote control on for target.dll.i64 (pid 4242)
http://127.0.0.1:8174
It then checks it is talking to the right database, and counts:
Agent: target.dll has 812 functions. That IDA wasn't serving SQL, so I turned on its remote control and started an HTTP server on localhost.
idasql remote off 4242andidasql http stop 4242undo both.
The whole exchange takes about 30 seconds. The agent switches remote control on only for the database you asked about.
Prefer to point the agent yourself? The classic way works as it always has. In IDA's
idasql> prompt type .http start. IDA prints IDASQL HTTP server: http://127.0.0.1:8174 and
copies the connection instructions to the clipboard. Then:
/idasql:connect Let's work with this database: http://127.0.0.1:8174
Run .pin http once, and that database starts its server every time it opens.
- No SQL to learn: you write plain-language prompts, and the agent writes the queries.
- From binary back to source code: take a compiled binary back to clean, fully recompilable source code. The agent recovers every name, type, struct and prototype, then writes out source you can build again. See real projects rebuilt this way in vibe-re.
- Headless or inside IDA: the agent runs IDA in the background, or finds the IDA you have open and connects. You can also point it yourself with
/idasql:connect. - Clickable answers: function names are
ida://links that jump IDA to the code (IDA 9.4+). - 70+ live tables, read and write: functions, xrefs, strings, types, pseudocode, ctree, microcode, comments, renames, prototypes and patches. Nothing is exported or indexed first.
- IDAPython on request: the agent writes a script and runs it in the IDA you name.
- MCP server built in:
idasql -s db.i64 --mcp, or.mcp startinside IDA, and any MCP (Model Context Protocol) client gets anidasql_querytool. - Pin it and forget it: run
.pin http(or.pin mcp) once in IDA, and that database serves every time it opens, so your agent connects with no step at all. - Many databases at once: compare, cross-reference, and move annotations between them.
- Agent-ready: 16 skills for Claude Code, Copilot CLI and Codex, plus HTTP and MCP servers for everything else. To see what agents build with it, check out vibe-re, a gallery of agentic "vibe" reverse engineering.
- A safety net:
idasql rescuesaves the work of a frozen IDA, andidasql closeends one cleanly. - IDA 9.2 to 9.5 on Windows, Linux and macOS, x64 and ARM64.
| You may have heard | Actually |
|---|---|
| "I have to learn SQL." | No. You ask in plain language, and the agent writes the SQL. |
| "It only reads." | It writes too: renames, comments, types, structs, prototypes, patches, bookmarks, and undo checkpoints. |
| "It exports the database first." | No. Every query runs live against the open database, and every change shows in IDA at once. |
| "It replaces IDA." | It runs on your IDA Pro (9.2 to 9.5) and your license. The pseudocode tables need the Hex-Rays decompiler. |
| "It only works headless" (or "only in the GUI"). | Either, or both, with many databases at once. |
| "It's for Claude only." | Any agent: the skills ship for Claude Code, Copilot CLI and Codex, and HTTP and MCP serve the rest. |
| "It's only for malware." | Any binary IDA opens: drivers, firmware, games, libraries. |
| "The agent will make things up." | Its answers come from queries on your database, and on IDA 9.4+ links let you click through to check each one in IDA. |
| "My binary goes to the cloud." | idasql runs on your machine and its servers listen on localhost only. Your agent's model sees what the agent reads, so pick your agent accordingly. |
-
Download the release for your IDA version from Releases (
ida92,ida93,ida94orida95; builds are IDA-version-specific):Archive What's in it *-plugin.zipThe HCLI-compatible plugin (Windows x64, Linux x64, macOS ARM64) *-cli.zipThe idasqlcommand line, by host architecture*-arm64-plugin.zipARM64 plugins, installed by hand: Linux (IDA 9.3+) and Windows (IDA 9.4+) -
Put
idasqlnext to IDA, with IDA's directory on your PATH, then check it withidasql --version. -
Add the skills to your agent. In Claude Code or Copilot CLI:
/plugin marketplace add allthingsida/idasql-skillsThen install
idasqlfrom that marketplace. For Codex, see idasql-skills.
$ idasql instances
PID HOST IDA SQL REMOTE STATE DATABASE
4242 gui 9.5 http://127.0.0.1:8174 on ok target.dll.i64
5150 gui 9.5 - off ok malware.exe.i64
6001 idalib 9.5 http://127.0.0.1:8123 - ok firmware.bin.i64
Every IDA on the machine is listed: GUI windows, idasql -s sessions, even IDAs without idasql
(9.4+). When one serves no SQL, the agent starts its server, using the same verbs you type in IDA:
| In IDA | From any terminal | From SQL |
|---|---|---|
.http start / .http stop |
idasql http start <db|pid> / stop |
UPDATE ida_instances SET start = 'http' WHERE pid = N |
.mcp start / .mcp stop |
idasql mcp start <db|pid> / stop |
UPDATE ida_instances SET start = 'mcp' WHERE pid = N |
.remote on / .remote off |
idasql remote on <db|pid> / off |
UPDATE ida_instances SET remote = 1 WHERE pid = N |
Remote control is off by default, and set per database. Every request appears in that IDA's
Output window with the process that asked. Servers bind to localhost, and no token is ever
handed out. Prefer to connect by hand? .http start and a pasted URL work just as before.
Take a compiled binary and get back clean, fully recompilable source code: real names, real types, real structs, and code that builds again.
/idasql:re-source Start at the entry point and recover this binary bottom-up: name every function, rename locals, recover the structs behind the pointer arithmetic, and fix the prototypes.
/idasql:types Recover the struct this function fills in, apply it, and show me the pseudocode again.
The agent works through the program one function at a time:
- it decompiles the function;
- it renames the function and its locals, and comments it;
- it turns
*(DWORD *)(a1 + 0x10)patterns into real structs and enums; - it applies the types and prototypes, then decompiles again to check that the code reads cleanly.
Everything is saved in your IDA database as it goes. Once no sub_* is left, the agent writes the
program out as clean source files, with headers for the recovered types, and builds it again.
vibe-re collects real binaries rebuilt this way: malware
command-and-control (C2) code, kernel drivers and their typed client libraries.
SELECT f.name, dl.uri FROM funcs f
JOIN deep_links dl ON dl.resource = 'functions' AND dl.addr = f.addr AND dl.view = 'pseudocode'
ORDER BY f.size DESC LIMIT 5;The agent writes these as [sub_401130](ida:///sample.exe.i64/functions?rva=0x1130&view=pseudocode&idb_path=...).
Click one and IDA jumps to its pseudocode. If no IDA has that database open, it starts one on it.
Run idasql link register once to handle ida:// links. Links are for IDA windows; answers
about headless sessions use plain names.
You: write an IDAPython script that lists every call to CreateFileW, with the function each call sits in, and run it in the target database
The agent writes the script:
import idautils, idc, ida_funcs
target = idc.get_name_ea_simple("CreateFileW")
for xref in idautils.XrefsTo(target):
print(f"{xref.frm:#x} {ida_funcs.get_func_name(xref.frm) or '?'}")Then it runs the script inside that IDA through SQL:
UPDATE runtime_settings SET value = '1' WHERE key = 'enable_idapython';
INSERT INTO idapython_runs(code, show_output) VALUES ('<the script above>', 1);
SELECT success, output, error FROM idapython_runs ORDER BY id DESC LIMIT 1;Agent: CreateFileW is called from 6 places in 4 functions, listed with the address of each call. The same list is in that IDA's Output window. Running Python through idasql is off by default, so I turned it on for this session.
Every run is a row in idapython_runs, holding its output, success and error. A script file works
too (path instead of code), and a named sandbox keeps its globals between runs.
One /idasql:connect starts the session:
/idasql:connect Open and analyze all the drivers in my c:\malware folder.
After that, just talk. You don't need /idasql:connect again:
- "Which of them talk to the registry?"
- "How many strings do they have in common?"
- "Rename the dispatch routines in all of them, then save all databases and shut down IDASQL."
The agent runs IDA headless (idasql -s), analyses raw binaries itself, and works with every
database at once: compare them, cross-reference them, or move annotations between them.
Everything above is SQL under the hood, and you can type it yourself too:
$ idasql -s target.dll.i64 -q "SELECT name, size FROM funcs ORDER BY size DESC LIMIT 10"
$ idasql -s target.dll -q "SELECT * FROM binary" # a raw binary is analysed first
$ idasql -s target.dll.i64 -w -q "UPDATE funcs SET name = 'parse_config' WHERE addr = 0x401000"
$ idasql -s target.dll.i64 -i # an interactive REPL
Results print as tables, -f runs a SQL file, and -w saves the changes. Writes are plain SQL:
UPDATE funcs SET name = ..., INSERT INTO comments ..., UPDATE bytes SET value = ....
$ idasql rescue demo.i64
waiting for the IDA holding ...\demo.i64 to save ...
saved: ...\demo.rescued-20261110-111807.i64
$ idasql close demo.i64 --save
rescue saves a copy of a database whose IDA has frozen and won't come back: a script stuck in
an endless loop or sleep, a plugin blocked in a call, a hung wait. Your unsaved work comes out
from another terminal while IDA stays frozen. It reaches IDA windows and idasql -s sessions
alike, and --kill then ends the frozen IDA. An IDA that is only busy (a script still calling
IDA, analysis running) may crash if you rescue it, so wait for it, or use --kill.
close ends a healthy IDA, with --save or --discard.
Ways to run: CLI, plugin, HTTP, MCP, autostart, remote control
| Mode | Start it with |
|---|---|
| CLI | idasql -s db.i64 -q "...", -i (REPL), -f file.sql, -w (save on exit) |
| Inside IDA | Choose idasql in the command-line dropdown, then type SQL or .help |
| HTTP | idasql -s db.i64 --http 8080, or .http start in IDA |
| MCP | idasql -s db.i64 --mcp, or .mcp start in IDA. The client config is {"mcpServers": {"idasql": {"url": "http://127.0.0.1:<port>/sse"}}}, and the tool is idasql_query |
| Autostart | .pin http [bind] [port] in IDA: the server starts every time this database opens (.pin, .pin off http, .pin clear all) |
| Remote control | .remote on in IDA, or idasql remote on <db|pid>: other processes may then start and stop this database's servers |
Querying the HTTP server:
curl -X POST http://127.0.0.1:8080/query --data-binary "SELECT name FROM funcs LIMIT 5; SELECT COUNT(*) FROM strings;"The other endpoints are /status, /help, /cancel and /shutdown. Every /query returns one
envelope: success, statement_count, and results[], each with columns, rows, row_count,
elapsed_ms and error. Statements stop at the first failure unless you pass
?continue_on_error=1. A server bound beyond localhost should get --token.
Exit codes for idasql http|mcp start|stop and idasql remote: 0 done, 1 not found or refused,
2 usage, 3 timed out (IDA busy or frozen), 4 remote control is off. An idasql -s --http session
reports the URL it already serves. Other -s sessions can't be started from outside.
Deep links (IDA 9.4+)
deep_linksmakes anida://link for any function, address, string or type, in IDA's own Copy Link format plusidb_path, the database's full path. Function links open the decompiler, and the rest open disassembly.deep_link_parsetakes a link apart.idasql link open -- "<uri>"steers the running IDA that has the database open, preferring the link's exactidb_path. If none has it open, it starts IDA (.i64/.idb), waits for analysis (--timeout), and navigates. It won't start a second IDA when one doesn't answer or the database's.id0exists, unless you pass--launch-anyway.idasql link registerandidasql link unregistermanage the click handler. A handler it replaces, such as hcli's, is saved and restored. Runidasql link statusto see the current one.- Inside IDA,
SELECT ida_open_link(uri)navigates the current session. - Windows Terminal asks before opening custom schemes unless you add
"safeUriSchemes": ["ida"]to its settings.
Tables (70+)
Full columns, write rules and examples are in the schema catalog.
| Area | Tables |
|---|---|
| Code | funcs, func_at(), function_chunks, fchunks, blocks, instructions, instruction_operands, heads, data_items, segments, names, entries, function_frames, switch_tables, switch_cases |
| Graphs | xrefs, data_refs, call_graph(), shortest_path(), cfg_edges, dominators, post_dominators, loops, disasm_calls, disasm_loops |
| Data | strings, bytes (patch with UPDATE), byte_search, imports, fixups, mappings |
| Decompiler | pseudocode, ctree, ctree_lvars, ctree_call_args, ctree_labels, microcode, microcode_operands, decompiler_splits |
| Types | types, types_members, types_enum_values, types_func_args, local_types, applied_types, struct_member_xrefs, type_gaps |
| Annotations | comments, bookmarks, notepad, hidden_ranges, checkpoints, undo_state, dirtree_folders, dirtree_entries |
| Search | grep (functions, labels, types and members by pattern) |
| IDA and sessions | binary, db_info, ida_info, ida_config, runtime_settings, problems, signatures, available_signatures, ida_instances, ui_actions, idapython_runs, deep_links, deep_link_parse, breakpoints, netnode_kv |
Scalars include decompile(addr), disasm_at(addr), disasm_func(addr) and ida_open_link(uri).
SELECT name, kind, printf('0x%X', addr) AS addr FROM grep WHERE pattern = 'Create%' LIMIT 20;Skills (16)
| Skill | For |
|---|---|
connect |
Sessions, servers, finding running IDAs, routing |
analysis |
Triage, security audit, crypto and network detection |
disassembly |
Functions, segments, instructions, blocks, CFG, signatures |
decompiler |
Pseudocode, ctree, locals, labels, cleanup |
xrefs |
Callers, callees, imports, call graphs |
data |
Strings, bytes, byte patterns |
types |
Structs, unions, enums, typedefs, applying types |
annotations |
Comments, renames, bookmarks, notepad, undo checkpoints |
re-source |
Recursive annotation, structure and type recovery |
grep |
Find anything by name pattern |
deep-links |
Clickable ida:// answers (9.4+) |
idapython |
Run IDAPython; query SQL from IDAPython |
ui-context |
What's on screen and selected in IDA |
debugger |
Breakpoints, patching, patch inventory |
storage |
Key-value storage in the database |
functions |
The SQL function catalog |
Try /idasql:analysis what does this binary do? or /idasql:xrefs who calls CreateFileW, and how are errors handled?
CLI help
$ idasql --help
idasql v0.0.20 - SQL interface to IDA databases
Copyright (c) 2024-2026 Elias Bachaalany
Usage: idasql -s <file> [-q <query>] [-f <file>] [-i] [--export <file>]
Options:
-s <file> IDA database (.idb/.i64) OR raw binary (.exe/.dll/firmware/etc.)
- raw binaries trigger fresh idalib analysis and string-list rebuild
- legacy 32-bit .idb files upgrade to .i64 and require an explicit reopen
--token <token> Auth token for HTTP server mode (MCP has no auth)
-q <sql> Execute SQL query or semicolon-separated script
-f <file> Execute SQL from file
-i Interactive REPL mode
-w, --write Save database on exit (persist changes)
--export <file> Export tables to SQL file (local mode only)
--export-tables=X Tables to export: * (all, default) or table1,table2,...
--http [port] Start HTTP REST server (default: 8080, local mode only)
--bind <addr> Bind address for HTTP/MCP server (default: 127.0.0.1)
--mcp [port] Start MCP server (default: random port, use in -i mode)
Or use .mcp start in interactive mode
-h, --help Show this help
--version Show version
Commands (no database is opened; `idasql <command> --help` for details):
link <verb> ida:// deep links: open | status | register | unregister | source | config
instances [--json] running IDA instances and the databases they have open
rescue <db> [--kill] save a copy of the database a stuck IDA has open
close <db|url> end an IDA cleanly: --save or --discard
http start|stop <db|pid> start or stop another IDA's HTTP server (needs remote on there)
mcp start|stop <db|pid> the same for its MCP server
remote on|off <db|pid> let other processes start/stop that IDA's servers
Examples:
idasql -s test.i64 -q "SELECT name, size FROM funcs LIMIT 10"
idasql -s test.i64 -f queries.sql
idasql -s test.i64 -i
idasql -s test.i64 --export dump.sql
idasql -s test.i64 --http 8080
idasql -s sample.exe --http # raw PE: idalib auto-analyzes, then serves SQL (default port 8080)
idasql -s firmware.bin -q "SELECT * FROM binary"
idasql -s test.i64 --mcp 9000
Legacy 32-bit .idb files are upgraded to a sibling .i64: idasql exits with code 3 and prints
{"status":"upgraded","reopen_with":...}. Reopen with -s <reopen_with>.
Building from source
Prerequisites
- CMake 3.20+
- C++20 compiler
- IDA SDK 9.0+ (set
IDASDKenvironment variable)
cmake -S . -B build -DIDASQL_WITH_MCP=ON -DIDASQL_BUILD_EXAMPLES=OFF
cmake --build build --config ReleaseUseful CMake switches:
| Switch | Default | Description |
|---|---|---|
IDASQL_WITH_MCP |
ON |
Build MCP server support via fastmcpp. Disable for a smaller/offline build or when you do not need --mcp / .mcp. |
IDASQL_BUILD_CLI |
ON |
Build the standalone idasql command-line tool. |
IDASQL_BUILD_PLUGIN |
ON |
Build the IDA plugin. |
IDASQL_BUILD_EXAMPLES |
ON |
Build the example programs under examples/. |
IDASQL_TARGET_IDA_VERSION |
>=9.0 |
IDA version or range written to the generated plugin manifest. Release packaging sets an exact version such as 9.3. |
IDASQL_IDA_CMAKE_GIT_TAG |
Reviewed lock SHA | Standalone ida-cmake revision from .github/release-revisions.json. |
IDASQL_LIBXSQL_GIT_TAG |
Reviewed lock SHA | Standalone libxsql revision from the same lock when no parent or sibling target is available. |
IDASQL_FASTMCPP_GIT_TAG |
Reviewed lock SHA | fastmcpp revision from the same lock when MCP support fetches it. |
Notes:
- Hex-Rays support is always compiled in and detected at runtime. If Hex-Rays is unavailable, decompiler-backed tables/functions are simply not registered.
- HTTP REST support is always compiled in; use
--httpfrom the CLI or.http startfrom the REPL/plugin CLI. - IDAPython SQL execution is compiled in but disabled by default at runtime. Enable it per session with
UPDATE runtime_settings SET value='1' WHERE key='enable_idapython';. IDASQL_WITH_MCP=ONfetchesfastmcpp;OFFremoves MCP support and the--mcp/.mcpcommands.XSQL_WITH_THINCLIENTis forcedON, andHTTPLIB_USE_OPENSSL_IF_AVAILABLEis forcedOFFbecause IDASQL uses local plain HTTP.
IDASQL is part of a family of tools that expose different binary-analysis and
debug-information platforms through the same SQL surface, all built on the
shared libxsql virtual-table framework. A
query you learn against one tool largely carries over to the others: the
same SELECT name, size FROM funcs ORDER BY size DESC LIMIT 10 runs everywhere.
Reverse-engineering platforms
Debug info & compiler data
- pdbsql: Windows PDB symbol files as SQL.
- dwarfsql: DWARF debug information as SQL.
- clangsql: Clang AST as SQL.
Core
- libxsql: the C++ SQLite virtual-table framework every tool above is built on.
-
libxsql - Header-only C++17 library for exposing C++ data structures as SQLite virtual tables. Provides the fluent builder API for defining tables, constraint pushdown, and HTTP thin-client support.
-
fastmcpp - Optional MCP server implementation used when building with
-DIDASQL_WITH_MCP=ON. -
ida-hcli (MIT, Hex-Rays SA) - the
idasql linkIPC client, launch and URL-handler behavior is reimplemented in C++ from hcli's; idasql does not need hcli installed. Its MIT notice is in THIRD-PARTY-NOTICES.md, also shipped in every release archive.
IDASQL is free to use, including commercially. If it saves you time in IDA, a donation funds the next release, and a star helps other reverse engineers find it.
Bug reports with a database and the exact query are just as valuable. Most fixes in this release started that way.
In short: you may read, build, evaluate, benchmark, package, and use unmodified idasql, including commercially, if you preserve notices and follow the license terms. You may fork or patch it to prepare bug fixes, optimizations, features, tests, or documentation improvements for contribution back within the license's contribution-purpose rules.
You may not maintain a divergent private fork, port, rebrand, clone, API-compatible replacement, competing implementation, or use idasql as AI input to recreate or improve a derivative implementation without prior written permission from Elias Bachaalany. Independent implementations that are not copied from, materially derived from, or substantially informed by idasql in the license's defined sense are not prohibited.
Permission requests: open a GitHub issue at allthingsida/idasql/issues.
If idasql materially informs a distributed project, preserve the human origin: credit idasql and Elias Bachaalany visibly in your README/docs and in About/credits UI when applicable. The license includes an examples/FAQ section for common allowed and permission-required uses. Third-party dependencies (libxsql, the IDA SDK, and their transitive dependencies) remain under their own licenses.
See the full Human-Origin Source License v1.0.
