Chelis is a numerical computing language for code that agents write and people supervise. Tensors carry named dimensions and precision in their type, and a proof stack checks the properties you state, so a shape error is rejected before anything runs: where numpy would broadcast three weights against a 3 by 1 return into a 3 by 3 grid and sum it into a plausible wrong number, chelis check rejects the multiply. The compiler checks shapes, precision, effects and ownership, `chelis check` answers in JSON with the error kind and source span, `chelis tide mcp` exposes check, eval, prove and structural edits to an agent, and `chelis prove` checks declared properties by type checking, an SMT solver or seeded sampling.
Be the first to rate Chelis. Your feedback helps other users decide.
Run chelis check and get shape, precision, effect and ownership errors as JSON with source spans.
Declare a property and let chelis prove check it by types, SMT or seeded sampling, naming the method used.
Expose check, eval, prove and structural edits over MCP with chelis tide mcp.
Similar tools in the same category — compare side-by-side
paste-preview is a Claude Code mod for the images you paste. In a terminal a pasted image is just [I...
what-would-it-cost is a Claude Code cost tracker: what your Pro or Max subscription would cost at AP...
Thyme is a native reverse-engineering IDE for Android and native binaries on macOS, Linux and Window...
Guides and comparisons about code tools
Nine launches that draw the line around an agent: policy proxies, guard hooks, semantic permissions, sandboxes, worktrees and a proof checker.
Four Show HN launches around voice and ops: voice planning for agents, a live game coach, CI for AI phone agents and a self-hosted MCP platform on Kubernetes.
Ten Show HN launches that answer the same question differently: where does what an agent learned live, who owns it, and who can read it back?
Ten recent launches for running agents where they can do less harm: virtual machines, air-gapped sandboxes, sequencers, and control planes you host yourself.
Ten launches aimed at the same bottleneck: agents re-read their own history, and every re-read is billed. Compression, code mode, prices and measurement.