cd /news/ai-agents/ramanujan-multi-model-agent-for-rese… · home topics ai-agents article
[ARTICLE · art-126679] src=github.com ↗ pub= topic=ai-agents verified=true sentiment=· neutral

Ramanujan – Multi-Model Agent for Research in Computational Mathematics

Developer AniketWathore released Ramanujan, a free terminal-based multi-model agentic workbench for research in computational mathematics, available on GitHub. Ramanujan spawns N parallel subagents across any OpenAI-compatible providers, runs each in an isolated worktree with local journal, budgets, and stall detection, and verifies results deterministically using SymPy and Z3 with independent double-verification. The tool requires macOS or Linux, Node.js 22+, npm, Python 3.12+ (SymPy, Z3, Pydantic), and at least one provider API key, and is still under active development.

read4 min views17 publishedSep 11, 2026
Ramanujan – Multi-Model Agent for Research in Computational Mathematics
Image: Michielbdejong (auto-discovered)

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.

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.

  • 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 explicitmodel-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.

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.

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.

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.

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.

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)

TypeScript TUI agent on the pi substrate (vendored) + frozen Python math engine spawned as a subprocess. The agent renders; only the engine computes.

git clone https://github.com/AniketWathore/Ramanujan.git
cd Ramanujan

./scripts/install.sh

Fresh machine, no checkout handy? The script is self-contained — prerequisites are just Node 22+, npm, and Python 3.12+.

Run ramanujan with an empty config — the setup wizard starts automatically:

  1. Pick a provider from the list.
  2. Paste your API key.
  3. Pick your main model from the live list.
  4. Build your computational pool (reuse providers or add new ones, multi-pick models).
  5. Save it as the default preset and confirm — you land on the home screen.
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
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 vspi ) 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 withRAMANUJAN_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
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
  • 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 attools/obscura/bin/obscura for literature web collection (papers, books, websites, blogs, articles, discussions as text).
  • SymPy andZ3 — the deterministic math backend.
  • Srinivasa Ramanujan — the name, and the standard.

Distributed under the MIT License. See LICENSE for more information.

── more in #ai-agents 4 stories · sorted by recency
── more on @ramanujan 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/ramanujan-multi-mode…] indexed:0 read:4min 2026-09-11 ·