Agentic Studio

A privacy-first, extensible AI coding agent that runs entirely on your own machine. Agentic Studio connects a local LLM (via Ollama) to your codebase, your terminal, your databases, and your documents — with every potentially dangerous action gated behind an explicit human approval, and nothing ever leaving your laptop.

Built for indie developers who want a Cursor/Claude Code–style workflow without sending their code or their data to the cloud.

Screenshots

Landing page	New chat	Multi-mode canvas
Local stats pulled straight from the codebase (tool count, test count, live states)	Quick-start prompts for exploring the project, running commands, querying the database, or generating documents	The code editor canvas open beside the chat, rendered instead of a raw code block

What it can do

• Codebase-aware chat — read files, list directories, and grep across your project so the agent can answer questions and make changes with real context.
• Safe file edits & terminal access — write_local_file, edit_local_file, and run_terminal_command all pause for your explicit approval before running anything, via a single-use, tamper-checked approval token.
• Database introspection — connect to a local SQLite file or a Postgres database and let the agent read the schema before writing queries.
• Real document generation — ask for an actual .docx, .xlsx, or .pptx file and get one saved to workspace/output/, downloadable straight from the chat.
• A multi-mode canvas — code editor, live HTML preview, rich-text documents, Mermaid diagrams, interactive spreadsheets, and charts, all rendered in a side panel instead of dumped as raw text in the chat.
• MCP support — connect any Model Context Protocol server from Settings and its tools become available to the agent immediately (always approval-gated, since a third-party tool's behavior can't be known in advance).
• Vision — attach images and a local vision sub-model describes them for the main model to reason about.

Requirements

• Node.js 20 or newer

• Ollama installed and running locally

• Two Ollama models pulled (or created) locally:

  • a main chat/tool-calling model (default tag: ornith:9b)
  • a smaller helper model used for vision descriptions and conversation-summarization (default tag: qwen3.5:4b)

  These default names are whatever models you've set up locally — if you're pulling stock models from the Ollama registry instead, pull something with solid tool-calling support (e.g. ollama pull qwen2.5:7b) and point Settings at it; see "Configuration" below.

Getting started

npm install
npm run dev

Then open http://localhost:3000 and click Enter Studio.

On first launch, open Settings in the Studio UI and confirm:

• Workspace root — the local folder the agent's file/terminal/database tools operate inside. Defaults to this app's own folder; point it at whatever project you want the agent to work on.
• Chat model / Sub-agent model — the Ollama model tags to use (see Requirements above).
• Ollama URL — defaults to http://127.0.0.1:11434; change it if Ollama runs elsewhere.
• MCP servers — optionally add external MCP servers (stdio-launched, e.g. npx -y @modelcontextprotocol/server-filesystem <path>) to extend the agent's tools.

If Ollama isn't running or the configured model isn't pulled, the chat will surface a clear error rather than hanging silently.

Local data & privacy

Everything the app stores lives on your machine, next to the project, and is git-ignored:

• .agentic-chats.db — chat history (SQLite)
• .agentic-settings.json — your workspace path, model choices, and any configured MCP servers (which may include API keys/tokens for third-party MCP servers)
• .agentic-memory.json — preferences you've asked the agent to remember
• workspace/output/ — generated .docx / .xlsx / .pptx files

Nothing here calls out to a cloud AI provider — inference happens through your local Ollama instance.

Development

npm run lint   # ESLint
npm test       # Vitest — unit tests for the security-critical modules (approval tokens, the
               # path-traversal guard, origin checking) and the chat storage layer

Known issues

• image-size (a dependency of pptxgenjs) has two open, unpatched denial-of-service advisories (GHSA-w3rx-r6r6-pgpr, GHSA-5p2g-fcmc-qvqq) affecting its ICNS/JXL/HEIF parsers — a specially crafted image can hang the process in an infinite loop. No patched release exists upstream as of this writing (checked directly against the advisories, not just npm audit's summary — every published version, including the latest, is listed as affected). npm audit fix --force's suggested fix (downgrading pptxgenjs to 1.1.5) is not a real fix and is not applied here. Accepted as a known risk, for two independently-confirmed reasons: (1) generate_presentation in this codebase never calls pptxgenjs's image-embedding API, and (2) inspecting pptxgenjs's own installed bundle (node_modules/pptxgenjs/dist/pptxgen.cjs.js) shows image-size is dead weight there too — the only reference to it is inside a comment block explicitly marked FIXME: TODO: currently unused; the only dependency pptxgenjs actually require()s at runtime is jszip. So the vulnerable parsers aren't reachable through any path in this app, nor through pptxgenjs itself as currently published. (We also confirmed there's no clean way to silence this in npm audit's report: the advisories have no version floor, so even a decade-old image-size release without ICNS/JXL/HEIF support at all still gets flagged. The only way to make the audit report 0 here would be aliasing image-size to an empty stub package via npm overrides — deliberately decided against, since a naive stand-in package can introduce its own vulnerabilities, as tested, and a proper zero-dependency stub adds fragility if pptxgenjs ever starts really using it.) Re-evaluate if image embedding is ever added to generate_presentation, or periodically check npm audit for an upstream fix.
• xlsx is pinned to a tarball from the SheetJS CDN (https://cdn.sheetjs.com/...) rather than the npm registry, because the npm-published xlsx package has been stuck on an old, vulnerable version (0.18.5) for a long time — see SheetJS's own installation docs. This is SheetJS's own documented recommendation, not a workaround.

Project structure

src/app/
  page.tsx                     Landing page
  studio/page.tsx               Main chat + canvas UI
  api/agent/route.ts            Orchestration loop (streams to/from Ollama, runs tools)
  api/agent/tools/               Tool implementations, settings, approvals, MCP client, chat storage
  api/agent/tools/execute/       Endpoint that runs an approved dangerous tool call
  api/chats/                     Chat list CRUD
  api/download/                  Serves generated files from workspace/output/
  api/settings/                  Reads/writes app settings

License

MIT — see LICENSE.