Retrieved article excerpt
Open article · Retrieved 2026-09-11T08:23:14.958856+00:00
Ramanujan Multi Model Agentic Workbench for Research in Computational Mathematics Ramanujan is a terminal based multi model agentic workbench for research in computational mathematics. It is a free tool to help with research in maths. You can chat normally and when you want to dig deeper it structures your problem, surveys the literature, runs parallel subagents and brings back a consolidated report. Everything runs locally and every verdict is computed deterministically. System Architecture Disclaimer: This project is still under active development. If you find any bugs, issues, or have suggestions, please open an issue on the GitHub Issues page . Features Multi-model parallel subagents — spawn N subagents on different models at once, with live progress and contradiction flagging. Isolated worktrees — each run gets its own worktree with local journal, budgets, and stall detection. Multiple providers & models — mix any OpenAI-compatible providers, multi-pick models, save them as presets. Verified literature survey — every source carries a real URL or an explicit model-memory, unverified tag. Deterministic checking — SymPy + Z3 kill-check with independent double-verification; cross-provider panel only advises where no check applies. Checkpoints & consolidated report — confirm / revise at every stage, then get a plain-language report with technical appendix. Workflow 1. First Launch — Setup in the TUI Open ramanujan with an empty config → provider picker → API key → main-model picker → computational pool (existing or new providers, multi-pick models) → save preset → confirm → home screen. Zero commands. 2. Research — Just Ask Type a problem in chat ("Goldbach's Conjecture: Every even integer >2 is sum of two primes — prove it"). The agent structures it, shows the spec, and waits for your confirm / revise / typed feedback before the next stage. 3. Literature — Verify, Then Continue Sources, data, and related work land in a text store; you see the important entries, verify them, confirm-and-continue or type what to change. 4. Computational — Parallel Subagents Pick the preset (or providers/models manually) and the headcount. Subagents work in parallel with live progress, shared orchestration files, cross-help, and contradiction flagging — then one consolidated findings report. Requirements Requirement Details OS macOS (tested), Linux Runtime Node.js 22+ Package Manager npm Python 3.12+ (math engine: SymPy, Z3, Pydantic) LLM access At least one provider API key (any OpenAI-compatible endpoint) Tech Stack TypeScript TUI agent on the pi substrate (vendored) + frozen Python math engine spawned as a subprocess. The agent renders; only the engine computes. Installation # Clone git clone https://github.com/AniketWathore/Ramanujan.git cd Ramanujan # Install (builds the agent, installs the `ramanujan` bin globally, # puts `ramanujan-engine` on PATH — never touches any pi install) ./scripts/install.sh Fresh machine, no checkout handy? The script is self-contained — prerequisites are just Node 22+, npm, and Python 3.12+. First-Time Setup Run ramanujan with an empty config — the setup wizard starts automatically: Pick a provider from the list. Paste your API key. Pick your main model from the live list. Build your computational pool (reuse providers or add new ones, multi-pick models). Save it as the default preset and confirm — you land on the home screen. Usage — TUI Only ramanujan That's it. First launch opens the setup wizard; afterwards you land on the ASCII home + chatbox. Talk like a normal chatbot, or ask it to research something and follow the checkpoints. To re-run setup any time, remove the config and relaunch: mv ~ /.config/ramanujan/config.toml ~ /ramanujan-config.bak && ramanujan Troubleshooting Symptom Fix Chat footer shows a different model than the setup pick Re-run setup (command above) — the pick is now written once and never overwritten by pool providers; /model still switches anytime Setup text invisible / wrong colours Fixed — the wizard queries the real terminal background before the first screen; update with git pull + ./scripts/install.sh ramanujan opens pi, or pi opens Ramanujan Fixed — separate bins ( ramanujan vs pi ) and separate dirs ( ~/.ramanujan vs ~/.pi ); reinstall both cleanly and the collision is gone Setup keeps re-appearing Config isn't persisting — check ~/.config/ramanujan/config.toml exists ( 0600 ); skip once with RAMANUJAN_NO_SETUP=1 ramanujan No models listed for a provider The wizard falls back to the built-in catalog, then manual slug entry — any of the three works Verification Checklist uv run pytest -q # 153 passed — engine: journal, killcheck, stages, calibration uv run ruff check . # clean npm run test --prefix agent/packages/bridge # 20 passed npm run test --prefix agent/packages/config # 18 passed npm run test --prefix agent/packages/math-tools # 33 passed ./scripts/install-obscura.sh # minimal no-render obscura 45M+40M, tools/obscura/bin/obscura --version tools/obscura/bin/obscura fetch https://example.com --dump text # Example Domain uv run python -m ramanujan.obscura_client # is_available() true # Manual: # 1. Fresh config → `ramanujan` → wizard → home + chat, zero commands # 2. Research prompt → spec card → confirm → literature (arXiv+Scholar+websites via Obscura, per-category text) → confirm → worktrees → report # 3. `pi --version` still genuine; `~/.pi` untouched Acknowledgements pi agent (MIT) — TUI substrate, vendored verbatim except the documented fork diffs. Obscura (Apache-2.0) — minimal headless browser ( fetch --dump markdown , no-render) bundled at tools/obscura/bin/obscura for literature web collection (papers, books, websites, blogs, articles, discussions as text). SymPy and Z3 — the deterministic math backend. Srinivasa Ramanujan — the name, and the standard. License Distributed under the MIT License . See LICENSE for more information.