Calivi is a self-hosted chat interface for your own Ollama and OpenAI-compatible servers. Every message goes to the exact server and model you choose — no hidden router deciding for you.
One docker compose up · SQLite, no external services · nine languages
Two dropdowns pick the server and model. Streaming answers with reasoning, tool calls, and per-message attribution — visible, not hidden.
Here's a script that archives everything in the current directory and excludes its own output:
#!/usr/bin/env bash set -euo pipefail # timestamped archive of the current folder stamp=$(date +"%Y-%m-%d_%H-%M-%S") zip -r "backup_${stamp}.zip" . -x "backup_*.zip"
Run it inside the folder you want to archive; the new .zip lands right there.
// mockup with placeholder data — not a real conversation
One toggle offers the tool layer; the model decides what to call. Bundled web search comes built in, and any remote MCP server plugs in beside it. Read-only tools run freely, state-changing ones ask you first.
A green light means the server answered and its model list arrived. Red servers are hidden from the chat picker.
Read-only tools from remote MCP servers. Streamable HTTP and the older HTTP+SSE transport.
Per-model system (initial) prompts. Key = exact model name, 'default' = fallback.
# hot-reloaded — edit and the next message uses it
default: |
You are a careful, concise assistant. Answer the
question that was asked. Say when you are unsure.
"qwen3:32b": |
A reasoning model. Take multi-step problems slowly
and show the steps that matter.
Tool layer settings. enabled: master switch. max_iterations: agentic loop cap. MCP tools are switched on and off individually in the MCP tab.
# read on every request — no restart
enabled: true
max_iterations: 8
tools:
web_search:
enabled: true
num_results: 5
Everyone sees only their own chats. The first account is the super admin and cannot be demoted or deleted.
A self-hosted, multi-user chat app that connects to Ollama and OpenAI-compatible servers from one interface. Server and model selection is manual. Images and documents, a tool layer with web search and MCP servers, and approval before anything changes state.
FastAPI + React · self-hosted · github.com/orkun-soylu/calivi
// mockup with placeholder data — advances every 3s, click a tab to stop
A documentation or search tool returns the passages that matched one query — never a complete reference. A model can read a partial excerpt and confidently report that the part it did not receive does not exist. Calivi does not claim to fix that. It shows you what the model actually got.
Asked about Pydantic v2 validator modes, a model sent one broad query to Context7, got back an
excerpt in which the word wrap never appears, and reported that the mode does not
exist — while closing with “this is from the current documentation”. A re-worded query returns
the exact sentence it needed. The documentation was fine; one retrieval was treated as exhaustive.
We wrote a system-layer instruction telling the model to re-query when something looked uncovered, and measured it over five runs per arm. It was worse: the false claim went from 0/5 to 2/5, and one run exhausted the iteration cap without producing an answer. Telling a model to “say plainly when something is not covered” appears to invite the confident claim of absence that was the bug. It was not shipped.
The failure is intermittent — 0/5 without the guard — so it is variance, not a defect a prompt can remove. The model cannot be made reliable here, but the operator can be given what it saw: every tool call is clickable and opens the raw output, stored with the message so it survives a reload.
### model_validator with mode='after' for cross-field validation Source: https://github.com/pydantic/pydantic/blob/main/pydantic/functional_validators.py @model_validator(mode='after') def verify_square(self) -> Self: if self.width != self.height: raise ValueError('width and height do not match') return self
wrap// the panel is real; the excerpt is trimmed for the page
Choose the server and model from the top bar for every message. Only reachable servers appear — no auto-router second-guessing you.
Native Ollama and any OpenAI-compatible API — OpenRouter, LM Studio, vLLM, llama.cpp — side by side through one interface.
Answers stream as they generate; reasoning models show their thinking in a separate pane. Stop any response with Esc.
Paste images to vision models. Attach PDFs, docx or code — extracted as lossless text, not OCR — on any model.
A bundled SearXNG, plus any remote MCP server — Context7, GitHub, Exa. The model calls them on its own initiative, and you see which tool ran, right in the conversation.
Cookie sessions, admin and user roles, each person scoped to their own chats. Registration you can close anytime.
# clone and bring it up $ git clone https://github.com/orkun-soylu/calivi.git $ cd calivi $ docker compose up -d --build ✓ calivi running on http://localhost:8090
Backend, frontend and a bundled SearXNG. Needs only Docker & Compose — data lives in SQLite.
The first sign-up becomes the admin. Add your servers under Settings — Ollama by host, OpenAI-compatible by base URL.
Pick a server and model from the top bar and start chatting. Behind HTTPS? Set COOKIE_SECURE=true.