Interactive agent
impact-vision
Conversational analyst with all 52 tools, slash commands, a tool router that ranks the right tool for a free-text request, and multi-step playbooks for deal screening, LP reporting, compliance and portfolio review.
The toolkit
One engine.
One engine.
Six ways in.
A conversational agent, a CLI, a REST API, an MCP server, a Streamlit dashboard and a single-file web console — all backed by the same runtime, the same tools, the same permission model. Pick whichever fits how your team already works.
Interfaces
The agent runtime is identical across all of them — same tools, same skills, same guardrails.
CLI
impact-vision catalog · framework · dd
Seven subcommand groups plus three service commands. The catalog, framework and DD groups run fully offline — no API key needed to search 787 IRIS+ metrics or 122 DD questions.
MCP server
impact-vision serve-mcp
Expose the whole surface plus five read-only resources to Claude, Cursor, VS Code or any MCP client over stdio or SSE. Your existing AI workflow gains impact expertise.
REST API
/api/v1/*
FastAPI gateway with an OpenAPI explorer at /docs, ready to drop behind your fund's own tooling or portal. Token auth via environment variable for production.
Streamlit dashboard
streamlit run …/dashboard/app.py
Six tabs — company assessment, IRIS+, DD, framework scans, the ESG toolbox and portfolio analysis — for people who would rather click than type.
Web console
impact-vision serve-web
Chat UI, typed tool console and REST API on one port. Drag-and-drop deck upload, streaming replies, artifacts panel. Single self-contained HTML files — no build step, no CDN.
Quick start
From clone to first assessment.
Python 3.11+ and Git are the only prerequisites. The IRIS+ catalogue and DD checklist ship with the package, so the catalogue, framework and DD commands work before you configure any model at all.
git clone https://github.com/joejoe168168/impact-vision.git cd impact-vision python -m venv .venv # Windows PowerShell .venv\Scripts\Activate.ps1 # macOS / Linux source .venv/bin/activate pip install -e ".[dev]"
impact-vision setup # interactive wizard: # OpenRouter · Anthropic · OpenAI · Ollama # or run fully local: impact-vision ollama-setup --model llama3.2
# 787 IRIS+ metrics, offline impact-vision catalog search "climate" impact-vision catalog stats # 36+ frameworks impact-vision framework list impact-vision framework scan "company description" impact-vision framework xref OI4112 # 122 DD questions / 34 categories impact-vision dd categories impact-vision dd analyze "./memo.txt"
# conversational impact-vision # one-shot impact-vision -p "analyze deck.pdf and generate an HTML impact report" # service surfaces impact-vision serve-web --open impact-vision serve-mcp
Using it inside Claude, Cursor or VS Code? impact-vision serve-mcp exposes the full tool surface and five read-only resources over MCP, so your existing assistant gains IRIS+, SDG and regulatory expertise without leaving the editor.
52 agent tools
What the agent can actually do.
Every tool below is exposed identically through the interactive agent, the web chat UI, the REST API and the MCP server.
| Group | Tools | Coverage |
|---|---|---|
| Tool routing | 1 | impact_advisor ranks the most relevant tools for a free-text request and suggests multi-step playbooks. |
| Pre-screen & core assessment | 7 | Pitch-deck intake, IRIS+ catalogue search, SDG mapper, 5-Dimension assessment, gap analysis, metric recommender, data-quality score. |
| Due diligence & evidence | — | 122-question checklist analysis, NESTA evidence rating, follow-up question generation, structured Q&A scoring loop, AI-extraction review. |
| Frameworks & compliance | — | Multi-framework assessment across 20+ standards, cross-reference lookup, theory of change, greenwashing and anti-greenwashing screens, jurisdiction packs. |
| Climate, nature & valuation | — | GHG Scope 1/2 inventory, PCAF financed emissions, SBTi pathway, TNFD LEAP, NGFS scenarios, SROI, impact-adjusted IRR, IFVI/VBA, QALY welfare, LCA/LCSA. |
| Trust infrastructure | — | Stakeholder voice collection, verification workspace with finding lifecycle, HMAC-signed evidence graph, LP narrative from verified data only, exit-impact durability. |
| Reporting & portfolio | — | HTML/XLSX/CSV/JSON/PDF reports, IC memo, DD questionnaire, LP DDQ export, investee portal, batch portfolio analysis, natural-language portfolio query. |
| Consultant engagements | — | Proposal, data-room and training workflows for advisers running assessments on behalf of funds. |
Full per-tool reference, including flags and return shapes, lives in the repository README.
Deliverables
What comes out the other end.
Reports are generated locally as self-contained files. Nothing is uploaded, nothing phones home.
Impact report
HTML · PDFInteractive Plotly charts: 5-Dimension radar, SDG alignment scoring in official UN colours, sector opportunities and risks, methodology appendix. Light, dark and white-label themes.
IC memo
HTML · DOCXInvestment-committee-ready summary: the impact thesis, what is evidenced, what is asserted, the open questions and the recommended conditions.
DD report & questionnaire
HTML · XLSX · DOCXThe 122-question checklist as a working document, with responses, evidence levels and category scores tracked per company.
Investee portal
HTMLA page you can hand to the company: what you need from them, why it matters and what is still outstanding.
LP DDQ export
XLSX · CSVAnswers assembled from verified data only, mapped to the disclosure the LP actually asked about.
Portfolio roll-up
HTML · XLSX · JSONWeighted 5D scores, coverage rates, benchmark comparison and exit-impact durability across the fund.
Under the hood
Local by default,
Local by default,
auditable by design.
Assessments are stored in SQLite. Metric catalogues are JSON on disk. Reports are self-contained HTML. Evidence graphs are HMAC-signed with an audit trail, and the verification workspace tracks each finding through its lifecycle with threaded comments.
The engine is built on OpenHarness and works with OpenAI, Anthropic Claude, OpenRouter or a local Ollama model. Choose Ollama and no company data leaves your laptop.
Language
Python 3.11+
Installable SDK, importable from your own notebooks and pipelines.
Storage
SQLite + JSON
No database to provision. Portable, inspectable, diff-able.
Models
Bring your own
OpenAI, Claude, OpenRouter, Ollama. Switch with one command.
Licence
MIT
Use it in commercial fund work. Fork it. Ship it internally.
Open source
Read the code before you trust the score.
That is rather the point. Every mapping, threshold and weighting is in the repository, and every generated number traces back to the text that produced it.