Ollama Client

Local-first browser sidepanel for chatting with local and remote LLM providers, with private chat history, local RAG, image input, browser-context tools, and optional web search.

Quick links: Install · Docs · Provider setup · Architecture · Privacy · Issues

What It Does

Ollama Client gives you a browser-native chat workspace for local and bring-your-own remote models:

• Chat from the browser sidepanel with streaming responses and cancellation.
• Use verified built-ins for Ollama, LM Studio, and llama.cpp; add other OpenAI-compatible servers or Anthropic from Settings.
• Upload files and use local retrieval-augmented generation over your own content.
• Attach images for vision-capable models.
• Let tool-capable models read the current tab, selected text, open tabs, uploaded files, local memory, and optionally the live web when the prompt calls for it.
• Capture selected page text into chat with the selection-button overlay.
• Keep chat history, sessions, files, settings, and embeddings on your machine by default.
• Review all stored data, create a full backup, or wipe it from one Privacy screen.
• Organize chats with tags; edit messages in place or fork an alternate branch.
• Export, restore, print, and manage local conversation history.

Supported Providers

Provider	Default endpoint	Notes
Ollama	http://localhost:11434	Default fallback with tool calling and fullest local model-management support
LM Studio	http://localhost:1234/v1	OpenAI-compatible chat, embeddings, tool calling, and LM Studio model discovery
llama.cpp server	http://localhost:8000/v1	OpenAI-compatible server via llama-server
OpenAI-compatible	User configured	Add vLLM, LocalAI, KoboldCPP, or another compatible endpoint
Anthropic	https://api.anthropic.com/v1	Optional remote provider using the native Claude Messages API

Model routing uses saved model-to-provider mappings first. If a mapping is missing, the historical fallback is Ollama.

Local RAG

The RAG pipeline is browser-first and local-first:

1. Files, chat text, and live page context use one extension-owned chunker.
2. Chunks are embedded through provider-native support, a shared embedding model, or Ollama fallback.
3. Hybrid retrieval combines keyword and dense search.
4. Retrieved snippets are injected into the prompt context before generation.

Chat/session/message/file history is SQLite-only through sql.js, persisted into IndexedDB. Vector embeddings still live in IndexedDB through the embeddings storage layer.

Browser Context, Images, and Tools

Ollama Client supports two context paths:

• Manual context: select tabs, selected text, files, or images before sending.
• Model-requested context: tool-capable models can call tools during a response to inspect the current page, list/read open tabs, read permission-gated recently closed or synced-device sessions, search indexed files, search local chat memory, use the most recent selected text, or search the live web when web search is enabled.

Tool calls run inside the extension and are shown in the reasoning trace with status, inputs, sources, and trimmed output previews. They do not create extra chat-history rows; only the final answer and trace metadata are persisted.

Before sending, the composer context tray shows what the model can see and consolidates tab selection, files, screenshots, knowledge, and web controls. After an answer, page/tab context, local knowledge, and web results appear in one grouped Sources sheet.

Recently closed and synced-session tools are read-only, require optional permission, and honor never-read exclusions. Restoring a session is not exposed until model actions have a real interactive approval boundary.

Web search appears to tool-capable models as a single web_search tool. Backend choice stays in Settings -> Knowledge, with SearXNG for local/self-hosted search and Brave Search or Tavily through API keys. Search config is device-local, API keys are masked, snippets are capped, and returned titles/snippets are treated as untrusted text.

For local/private web search:

cd searxng
docker compose up -d

Then set the SearXNG endpoint to http://localhost:8080 in Settings -> Knowledge -> Web Search.

Image input is available only when the selected model resolves to vision-capable. Images are sent in the provider's native request format and stored locally with the conversation so previews reopen later.

Install

Chrome Web Store

1. Install from the Chrome Web Store.
2. Start at least one provider server.
3. Open extension settings, configure the provider URL, and select a model.
4. Start chatting from the sidepanel.

Local Development

git clone https://github.com/Shishir435/ollama-client.git
cd ollama-client
pnpm install
pnpm dev

Firefox development:

pnpm dev:firefox

Common Commands

pnpm dev                    # Chrome MV3 dev build
pnpm dev:firefox            # Firefox MV2 dev build
pnpm build                  # Chrome production build
pnpm build:firefox          # Firefox production build
pnpm package                # Zip Chrome build
pnpm package:firefox        # Zip Firefox build

pnpm typecheck
pnpm lint:check
pnpm test:run
pnpm verify                 # typecheck + lint + full test suite
pnpm verify:browser-smoke
pnpm verify:browser-automation

Before opening a PR, run:

pnpm verify

Architecture

The extension is built with WXT, React 19, TypeScript 5.9, Tailwind v4, and Biome.

Key paths:

• src/entrypoints/ - WXT entrypoints for sidepanel, options, background, content scripts, and print export.
• src/sidepanel/ - main chat shell.
• src/options/ - settings and configuration shell.
• src/background/ - runtime message dispatcher and handlers.
• src/features/ - feature-owned UI, hooks, RAG, stores, and workflows.
• src/components/forms/, src/components/layout/, src/components/settings/, src/components/feedback/, src/components/data-display/ - app-owned frontend primitives.
• src/components/ui/ - curated shadcn/Base UI primitives only.
• src/lib/providers/ - provider registry, factory, manager, and provider implementations.
• src/lib/repositories/chat-history.ts - chat-history facade backed by SQLite.
• src/lib/sqlite/ - sql.js database, schema, and migrations.
• src/lib/embeddings/ - chunking, embedding strategy, HNSW, keyword index, and vector storage.

Runtime flow:

1. Sidepanel sends a provider stream request through a runtime port.
2. Background dispatches to a provider handler.
3. ProviderFactory resolves the selected model's provider.
4. Provider streams tokens back through the port.
5. UI state updates and chat history is persisted locally.

Documentation

• Provider setup
• Context, images, and tools
• Provider capability matrix
• Architecture
• Keyboard shortcuts
• Changelog
• Privacy policy
• Contributing guide
• AI assistant guide

Search provider API references:

• SearXNG Search API
• Brave Search API
• Tavily Search API

Docs site source lives in docs/. The changelog and provider matrix docs pages are generated by pnpm docs:generate from root/source files, so do not hand-edit docs/src/content/docs/about/changelog.md or docs/src/content/docs/concepts/provider-matrix.md. The docs build output goes to docs/dist/ for Vercel. Generated extension locale metadata under public/_locales/ comes from src/locales/<lang>/translation.json; do not hand-edit generated locale files.

Privacy

Ollama Client does not include a built-in telemetry pipeline. Your privacy depends on the providers you configure:

• Local providers keep requests on your machine or local network.
• Remote providers receive the prompts, context, and files snippets you send to them.
• Remote search providers receive search queries when web search is enabled and selected.
• Chat history and RAG data are stored locally by default.

Do not expose local provider APIs publicly without authentication and network controls.

Contributing

Keep changes scoped, testable, and aligned with the existing feature boundaries. New chat-history work should go through src/lib/repositories/chat-history.ts; new provider work should update the provider registry, factory, manager defaults, tests, and provider docs.

See CONTRIBUTING.md and AGENTS.md for the full contributor workflow.

License

MIT License: LICENCE