gw

Grant Writer (gw)

Grant writing pipeline for Claude Code with multi-model consensus via claude-multi-model (shipped as a Python dependency — no API keys, no extra plugins, just uses your existing codex / gemini / cursor-agent CLI tools). Full proposal lifecycle from FOA analysis through writing, fact-checking, and peer review — targeting EU (Horizon Europe, ERC, MSCA) and Romanian (UEFISCDI, PNRR) funding agencies.

Markdown throughout — no LaTeX needed. Final Word/PDF conversion via /docx or /pdf skills at submission time.

Install

One-liner:

curl -fsSL https://raw.githubusercontent.com/stmailabs/gw/main/scripts/install.sh | bash

This installs:

• gw-skills Python tools in .venv/ (with claude-multi-model pulled in transitively, so mm-ask / mm-council / mm-detect become available via uv run)
• gw@gw plugin (this pipeline)
• Default config.yaml
• CLAUDE.md with CLI rules and skill table
• Probes for provider CLIs (codex, gemini, cursor-agent) and reports how many rate buckets are available

Verify with:

uv run gw-verify

Quick start

# Interactive mode — guides you through agency selection
claude '/gw'

# With explicit arguments
claude '/gw --agency horizon_ria --mechanism ria --foa path/to/foa.pdf'

# Resume from an existing proposal directory
claude '/gw --proposal-dir proposals/horizon_ria_ria_20260401_120000'

# Skip review for a quick draft
claude '/gw --skip-review'

Pipeline

 0.    Setup              →  environment, mm-detect probe, agency config
 0.5   Import Documents   →  auto-extract from PI's existing docs
 1.    FOA Analysis       →  parse funding opportunity, extract requirements
 1.5   Competitive        →  funded grants DB + literature + market analysis
       Landscape          →  mm-ask parallel research across 3 external models
 1.7   Deep Verification  →  SPA/JS inspection, transliteration, national registries
 2.    Aims Generation    →  iterative refinement + mm-ask adversarial critique
 3.    Literature         →  mm-ask citation discovery + CrossRef/S2 verification
 4.    Preliminary Data   →  assess PI's existing evidence
 5.    Proposal Writing   →  per-section atomic checkpoints + mm-ask pre-verification
 5.5   Risk & Feasibility →  mm-ask scenario analysis + risk matrix
 6.    Budget             →  person-months, equipment, travel, subcontracts
 7.    Supporting Docs    →  CVs, facilities, DMP, ethics, consortium
 8.    Compliance         →  word counts, required sections, structure validation
 8.5   Assembly           →  compile all sections into final/proposal.md
 8.7   Fact Check         →  mm-ask citations + external facts + claim-source alignment
 9.    Review             →  4-persona Claude panel (Agent tool) + 3-model mm-ask external track
 9.5   Resubmission       →  parse previous reviews, plan revisions
10.    Revision           →  weakness → phase mapping, surgical re-runs, re-review

Supported agencies

Template	Agency	Mechanism	Region	Budget Model
horizon_ria	Horizon Europe	RIA	EU	person-months
horizon_ia	Horizon Europe	IA	EU	person-months
erc	ERC	Starting / Consolidator / Advanced	EU	person-months
msca_postdoc	MSCA	Postdoctoral Fellowships	EU	person-months
msca_doctoral	MSCA	Doctoral Networks	EU	person-months
uefiscdi_pce	UEFISCDI	PCE (Exploratory Research)	Romania	monthly salary
uefiscdi_te	UEFISCDI	TE (Young Research Teams)	Romania	monthly salary
uefiscdi_pd	UEFISCDI	PD (Postdoctoral Research)	Romania	monthly salary
pnrr	PNRR	Component 9 (R&D Support)	Romania	monthly salary

Multi-model consensus via claude-multi-model

gw depends on claude-multi-model as a Python package. It exposes three CLIs — mm-ask, mm-council, and mm-detect — that dispatch prompts to external CLI tools in parallel and route each model to its native CLI:

• GPT / Codex family → codex exec (Codex/OpenAI OAuth — bucket 1)
• Gemini family → gemini -p (Google AI OAuth — bucket 2)
• Grok / Kimi / Composer / other → cursor-agent --model X (Cursor subscription — bucket 3)
• Claude family → refused by default (use the Claude Code Agent tool inside skills; the Agent tool is in-process and free)

With all three provider CLIs installed, a 3-model panel spreads across 3 independent rate buckets, so one hot bucket can't throttle the rest. Install whichever provider CLIs you have access to — the router adapts automatically and reports available buckets via mm-detect.

gw uses multi-model consensus in four load-bearing places: