Radicand

An ordinary calculator, and most programming languages, compute sqrt(12) as 3.4641016151377544 — an approximation that has already thrown information away. 1/3 + 1/6 becomes 0.4999999999999999. Fine for engineering estimates; noise if you are deriving, checking or teaching, where what you want is 2√3 and 1/2.

Radicand's position is in its name — the radicand is the number under the radical sign. It stays exact by default, giving numerics only when you ask, with precision chosen at evaluation time rather than fixed at storage time:

1/3 + 1/6      ->  1/2
sqrt(12)       ->  2*sqrt(3)
sqrt(-4)       ->  2*i
2^100          ->  1267650600228229401496703205376
N(pi, 60)      ->  3.14159265358979323846264338327950288419716939937510582097494
0.1 + 0.2      ->  0.30000000000000004

That last line is deliberate: a decimal in the input is what introduces inexactness, and it stays visible — approximate results carry a trailing dot.

Coverage:

• Algebra: expand((x+y)^5), factor(x^4+4), apart(1/(x^3-x), x), simplify, together
• Calculus: diff, integrate including definite integrals, limit, series
• Multivariable: gradient, hessian, jacobian, curl, laplacian, multiple integrals
• Differential equation solving
• Linear algebra: det, inv, eigenvals
• Number theory: factorint, gcd, factorials
• Plotting, plus a command-line REPL (radicand-cli)

A few engineering decisions are worth stating, because they determine how it feels to use:

• The kernel is a hash-consed DAG. Every distinct subexpression is stored once and referenced by a 4-byte Id, so structural equality is an integer comparison rather than a tree walk, and x + x is already 2*x at construction time. This is the SymEngine class of design.
• Offline structurally, not by promise. There is no HTTP client in the dependency tree, the front end contains no fetch, XHR or WebSocket, the CSP is default-src 'self', and KaTeX fonts are bundled locally. It is not that we pledge not to phone home; it has no capability to.
• The licensing is clean. MIT, and every package in the dependency tree is MIT or Apache-2.0 — no GMP, MPFR, Maxima, PARI or GiNaC. Writing the kernel from scratch instead of wrapping SymEngine was largely for this reason: you can embed it in anything, including closed-source commercial products. Most computer algebra systems cannot offer that.
• The web version is the same kernel compiled to WebAssembly and running in your browser, so radicand.doaipm.com is a full trial with nothing installed and nothing leaving your machine.

This one is ours, which makes it more important to be plain about the limits:

• It is early — currently v0.1.0, first released in August 2026. The feature range is real and the kernel carries 188 tests, but a 0.1 release means you will find uncovered cases and bugs. Cross-check anything that matters.
• It is not a Mathematica or MATLAB replacement. Their breadth — special functions, numerical methods, toolboxes, visualization, documentation ecosystems — is in another league. Radicand does exact symbolic computation; it is not a numerical simulation environment.
• No notebook workflow. There is nothing like Jupyter's interleaving of code, output and prose into a document. For reproducible analysis write-ups, use Jupyter.
• Plotting is basic — enough to see the shape of a function, not a publication figure tool.
• Documentation is mostly English so far, despite the landing page covering eight languages.

When it is the right choice: derivations that must stay exact, checking results from a textbook or paper, wanting a few-megabyte math tool that opens instantly with no network and no account, or embedding a cleanly licensed symbolic engine into your own software.