▶ Live demo: all 514 indicators over real Binance market data, computed live in your browser — live.wickra.org · zero backend, powered by
wickra-wasm.
A local market copilot: an LLM grounded in real order book, liquidation and funding microstructure — for R. install.packages("wickracopilot", repos = "https://wickra-lib.r-universe.dev") — over the C ABI via .Call, prebuilt library fetched on install.
R bindings for the wickra-copilot deterministic market-context core, over its C
ABI hub (.Call). Build a copilot from a spec JSON, drive it with command JSON,
read back the MarketContext — the same protocol as the CLI and every other
binding. Only the deterministic core is exposed; the LLM adapter is never
reachable over the C ABI, so the network and API key stay off this surface.
From r-universe:
install.packages("wickracopilot", repos = "https://wickra-lib.r-universe.dev")
The package's configure downloads the prebuilt C ABI library for this exact
version from the GitHub release and bundles it, so an ordinary install needs
nothing but a C toolchain (Rtools on Windows) for the thin .Call glue layer. To
build against a local checkout instead, point it at the header and library with
the environment variables below.
The package compiles against the wickra-copilot C ABI. Point the build at the
header and library with two environment variables (set by CI / the installer):
WKCOPILOT_INC — the directory holding wickra_copilot.h (i.e. bindings/c/include).WKCOPILOT_LIB — the directory holding the built shared library (i.e. the
Cargo target/release after cargo build -p wickra-copilot-c --release).At run time the loader finds the shared library via PATH (Windows) or
LD_LIBRARY_PATH / DYLD_LIBRARY_PATH (Linux/macOS).
cargo build -p wickra-copilot-c --release
WKCOPILOT_INC=../c/include WKCOPILOT_LIB=../../target/release R CMD INSTALL .
Rscript tests/run_tests.R
library(wickracopilot)
spec <- '{"symbols":["BTCUSDT"],"lookback":3,"facts":["price_move"]}'
copilot <- wkcopilot_new(spec)
feeds <- paste0(
'{"cmd":"build_context","feeds":{"BTCUSDT":{"symbol":"BTCUSDT","candles":[',
'{"ts":1,"open":100,"high":100,"low":100,"close":100,"volume":1},',
'{"ts":2,"open":97,"high":97,"low":97,"close":97,"volume":1},',
'{"ts":3,"open":94,"high":94,"low":94,"close":94,"volume":1}]}}}'
)
cat(wkcopilot_command(copilot, feeds), "\n")
cat(wkcopilot_version(), "\n")
| Function | Description |
|---|---|
wkcopilot_new(spec_json) |
Build a copilot from a spec JSON (errors on an invalid spec). |
wkcopilot_command(copilot, cmd_json) |
Apply a command JSON (set_spec, build_context, facts, query, reset, version) and return the response JSON. |
wkcopilot_version() |
The library version. |
The handle is an external pointer with a finalizer, so it is freed
automatically. Domain errors (a bad spec, an unknown command) come back in-band
as {"ok":false,"error":...}; only unusable arguments and caught panics raise.
Every binding forwards to the same data-driven Rust core, so what this one adds is
the call overhead of R's native .Call interface over the C ABI, not a different result. The core's throughput is
measured by the repository's benchmark suite and the nightly bench.yml run; the
numbers, the machine and how to reproduce them are in the repository
BENCHMARKS.md.
The full guide, the spec reference and the API documentation live in the main repository and the documentation site:
examples/r/
Wickra Copilot ships native bindings for Python, Node.js, WASM and Rust, plus a C ABI hub that any
C-capable language (C, C++, C#, Go, Java, R) links against — all forwarding to the
same data-driven, unsafe-forbidden Rust core.
Found a security issue? Please don't open a public issue. Report it privately
via the repository's Security tab ("Report a vulnerability") or email
support@wickra.org with a subject line starting [wickra security]. Full
policy: https://github.com/wickra-lib/wickra-copilot/blob/main/SECURITY.md.
Wickra Copilot is analysis software: it builds a deterministic market context and relays it to a language model of your choosing. It is provided "as is", without warranty of any kind. LLM output can be wrong and is not financial advice; the copilot only reports facts and places no orders. Trading carries risk of loss; review the code and use at your own discretion.
Licensed under either of Apache-2.0 or MIT at your option.