Interactive DuckDB shell implemented in Solod (So) and linked against libduckdb.

Experimental research code. SoloDuckDB is an AI assisted proof-of-concept to explore embedding DuckDB with Solod. It is not a supported project, not audited for security or correctness, and not suitable for production workloads. Prefer the official DuckDB CLI for anything that matters.

This research repository is self-contained and vendors the Solod compiler/stdlib as a git submodule (solod → solod-dev/solod). Builds always sync to the latest main on that repo (make runs scripts/update_solod.sh before translate; CI does the same).

CLI behavior is loosely aligned with the official DuckDB CLI (LTS):

• CLI overview
• Arguments
• Dot commands

% soloduck -c "SELECT version()"
┌-------------┐
│ "version"() │
│   varchar   │
├-------------┤
│ v1.5.2      │
└-------------┘

One-line installer (Linux/macOS, amd64/arm64)

curl -fsSL https://lmangani.github.io/soloduck/install | sh

Installs soloduck to ~/.soloduck/cli/latest/soloduck. Override latest with SOLODUCK_VERSION=<tag>.

git clone --recurse-submodules https://github.com/lmangani/soloduck.git
cd soloduck
make
./soloduck -version

If you cloned without submodules:

git submodule update --init --recursive

make fetches the latest Solod main into solod/ before so translate. To record that pointer in git: git add solod && git commit.

You need libduckdb on the machine (install)

make DUCK_PREFIX=/usr/local

make runs Solod so translate into gen/, then compiles and links soloduck next to the Makefile. Use make clean to remove generated C.

Show flags:

./soloduck -help

One-shot SQL (same idea as official duckdb -c):

./soloduck -batch -c "SELECT version();"
./soloduck :memory: 'SELECT 42 AS n'

• Interactive REPL (no -c and no SQL argument)
• -batch: read stdin as a script with no prompts (pipes and files).
• Output modes: -csv, -json, or .mode …. Default duckbox tries to mirror the official CLI layout

This compares the official DuckDB CLI against Solod + DuckDB’s C API in soloduck

Reference benchmark (python3 scripts/benchmark_init.py --runs 25 --warmup 2 on Apple Silicon M4;

Wall time per fresh process (startup + query + exit).
Timed runs per binary (after warmup): 25

Query: 'SELECT 1;'
  duckdb      mean    15.11 ms  σ   0.64 ms  min   14.19 ms  max   16.55 ms
  soloduck    mean    14.54 ms  σ   0.62 ms  min   13.45 ms  max   15.85 ms
  ratio (soloduck / duckdb mean wall time): 0.96x

Query: 'SELECT version();'
  duckdb      mean    15.10 ms  σ   0.68 ms  min   13.92 ms  max   16.74 ms
  soloduck    mean    14.26 ms  σ   0.57 ms  min   13.44 ms  max   15.17 ms
  ratio (soloduck / duckdb mean wall time): 0.94x

This compares Go runtime + database/sql + duckdb-go against Solod + DuckDB’s C API in soloduck

Reference benchmark (python3 scripts/benchmark_init_go.py --runs 25 --warmup 2 on Apple Silicon M4;

Wall time per fresh process (startup + query + exit).
Timed runs per binary (after warmup): 25
Compare: soloduck (Solod + libduckdb C API) vs benchmark_go_once (Go + database/sql + duckdb-go / CGO).

Query: 'SELECT 1;'
  duckdb-go     mean    17.01 ms  σ   0.75 ms  min   15.91 ms  max   18.62 ms
  soloduck      mean    14.55 ms  σ   0.48 ms  min   13.77 ms  max   15.40 ms
  ratio (soloduck / duckdb-go mean wall time): 0.86x

Query: 'SELECT version();'
  duckdb-go     mean    17.18 ms  σ   1.28 ms  min   15.71 ms  max   21.64 ms
  soloduck      mean    14.83 ms  σ   0.78 ms  min   13.79 ms  max   16.74 ms
  ratio (soloduck / duckdb-go mean wall time): 0.86x

go.mod uses replace solod.dev => ./solod so imports resolve to the submodule checkout.