An astronomical engine AI agents can compute with.
A clean-room ephemeris and Vedic computation engine, written in Rust from the primary sources. Agents call it over MCP instead of recalling planetary positions from training data, and every formula is measured against NASA JPL Horizons.
01 · Where the AI sits
How an AI agent uses Vedaksha
The agent
Asks for a position
Any MCP client discovers the tools and asks for a chart, a transit or a dasha period instead of guessing one.
Vedaksha MCP server
Seventeen tools
Served over stdio or over HTTP with a bearer token, from a tool catalog that is tested against the code so the two cannot drift.
The engine
The astronomy
JPL DE440/441 kernels, or the analytical VSOP87A and ELP/MPP02 theories, measured against JPL Horizons.
Published accuracy
A mean longitude error of 0.103 arcseconds against JPL Horizons on the kernel path, for 1900 to 2025, with a public parity harness.
Vedic layer
Nakshatras, five dasha systems, sixteen vargas, Shadbala, Ashtakavarga, chara karakas and gochara.
Western layer
Ten house systems, eleven sidereal ayanamshas, aspects, dignities, synastry and composite charts.
Graph output
Charts emit Cypher, SurrealQL, JSON-LD and embedding text for retrieval.
Fast paths
Vectorised trigonometry on AVX2 and NEON, batch charts, parallel transit search and a no_std core.
Every runtime
Rust crates, a Python package that runs the Rust engine through wasmtime, a WebAssembly build for the browser, and an MCP server image.
02 · How agents built it
What agents wrote and what people decided
Agents drafted
Crate scaffolding, most of the test suite, docs, Python and WebAssembly bindings.
Humans designed
The clean-room reading of the JPL ephemeris and IAU specifications, the Rust API surface, the MCP tool contract and the source-available licensing model.
Gated on every change
Every formula is checked against JPL Horizons oracle data. Every public API signature and every MCP tool schema is reviewed before it ships.
03 · Stack
What it runs on
One Rust workspace, distributed to every runtime an agent or an app is likely to use.
- Language
- Rust, edition 2024
- Crates
- 9 in the workspace, 7 on crates.io
- Ephemeris
- JPL DE440/441 · VSOP87A · ELP/MPP02
- Distribution
- crates.io, PyPI, npm, GHCR, GitHub Releases
- Agent surface
- MCP over stdio or HTTP
- License
- BUSL-1.1, each version Apache 2.0 after four years