Built so you can crack it open.
Everything is a logged tool call. Every skill is a typed contract. Every screen is React under your control via the VS Code extension. If you've ever wanted to extend the planning platform you use, this is the one.
Architecture
One backend, governed by typed skills.
A planner's question becomes a sequence of typed skill calls — connect-to-data, build-model, run-scenario, query-results — each parameterised and logged. The backend is FastAPI in front of PandaPower. The frontend is React shipped two ways: as VS Code webviews under the ⚡ Synth Activity Bar, and as a per-tenant hosted web app behind synth-headless ingress on Azure Container Apps.
Skills
Typed contracts, not magic.
AgenticGrid ships as Claude Agent SDK skills (under .claude/skills/agenticgrid-*). Each is a markdown file with a fenced YAML frontmatter declaring its triggers and arguments. Skills are how the agent reaches the simulator and the regulatory corpus — and how you extend it. Add a new analysis type? Write a new skill.
Run it locally
VS Code extension, your data, your terms.
The extension shows up as the ⚡ Synth Activity Bar icon. From there: a primary sidebar with the Modules launcher, AgenticGrid scenario tree and All Loops, plus editor-area webview panels and a secondary-sidebar remote-agent chat. VS Code's own Claude Code panel handles the chat surface — Ruflo, Ollama, sidecars, secrets and remote provisioning all live in the Node host under vscode-extension/src/host/.
Deploy a tenant
One skill, one command.
The agenticgrid-deploy-webapp skill provisions a new tenant end-to-end: image push, Azure Files shares, bcrypt admin password, JWT signing secret, container-app provisioning, custom domain. No manual ARM templates.
Want to extend it?
We're hiring engineers who want to build governed AI products in domains that matter.
Talk to us