ochat (formerly ollama-chat)

Made with Claude to talk with non-Claude.

Named "ollama-chat" because it started that way. Now a multi-backend TUI client (Ollama, OpenAI-compatible, llama.cpp) — just ochat.

A setup-and-forget multi-backend TUI chat client. One screen, no sub-menus, plenty of creature comforts.

Why?

I wanted something like Open WebUI but in terminal. A chat interface without needing a full browser eating 8GB of RAM just to exist.

It started as a bare-bones wrapper, but grew into something I actually enjoy using daily — switchable personalities, auto-suggest, impersonate mode, context tracking, multiple config profiles. Still not Open WebUI, but not bare-bones either.

I also wanted a TUI as pleasant to use as Claude Code but for local Ollama models.

This project was entirely vibe-coded with Claude. It does exactly what I need.

Features

• Clean TUI with streaming responses
• Switchable personalities (system prompts)
• Multiple config profiles (easily switch between setups)
• Project-specific prompts (auto-loads agent.md/system.md from current directory)
• Slash commands (/help, /retry, /undo, /personality, /config, /impersonate, etc.)
• Keyboard shortcuts (Ctrl+C cascade, Escape to cancel generation, Ctrl+R retry, Ctrl+U undo, Ctrl+G generate, etc.)
• Auto-suggest: after each response, a short suggestion appears in the input (Tab to accept)
• Impersonate mode: LLM suggests what you'd say next (/imp long-form, /imps short)
• Context tracking with usage warnings and /compact to summarize conversation
• Generation stats (TTFT, tokens/s) in status bar and /stats
• Persistent configuration
• Pass-through for any Ollama model option (temperature, top_p, etc.) via config file
• Multi-backend support: Ollama (default), OpenAI-compatible (LM Studio, vLLM), llama.cpp server, or auto-detect

Non-features

• It's just a client. No plan to have it start Ollama by itself or llama-cpp-python or anything
• Made for Ollama first, but supports OpenAI-compatible servers (LM Studio, vLLM) and llama.cpp server as first-class backends
• No conversation persistence yet (could use agent.md generation as lightweight conversation memory in the future)
• No multi-model conversations
• No model templates (chat templates are handled by Ollama/your server)
• No multiline input (TextArea doesn't support suggesters, and I prefer a single-line input with autocomplete over multiline without it — or making Claude reinvent the wheel and turning this codebase from "only Claude and God understand it" to "only God understands it")
• No RAG, no agents, no tools

Installation

git clone https://github.com/baptisterajaut/ollama-chat
cd ollama-chat

# Optional: symlink to your bin
ln -s "$(pwd)/ochat.sh" ~/bin/ochat

The virtual environment and dependencies are created automatically on first run.

Windows (untested)

git clone https://github.com/baptisterajaut/ollama-chat
cd ollama-chat
ochat.bat

Windows support is experimental and untested.

Usage

ochat                    # Start chatting (first run launches setup wizard)
ochat -C                 # Configure (host, model, personality)
ochat -m llama3.2        # Override model
ochat --new              # Create a new named config profile
ochat --use-config NAME  # Use a named config for this session
ochat --use-config NAME --as-default  # Switch to named config permanently
ochat -d                 # Enable debug logging
ochat --help             # Show options

Keyboard shortcuts

Key	Action
Ctrl+C	Cascade: clear input → cancel generation → double-press to quit
Ctrl+D	Quit
Ctrl+L	Clear chat (double-press to confirm)
Ctrl+R	Retry last response (double-press to confirm)
Ctrl+U	Undo last exchange (double-press to confirm)
Ctrl+G	Generate user message (ignored if input is non-empty)
Escape	Cancel generation
Tab	Accept suggestion / autocomplete command

Destructive shortcuts (clear, retry, undo) require a confirmation press within 2s. Config toggles (streaming, thinking) are command-only (/stream, /thinking) — they're settings, not runtime actions.

Slash commands

Command	Action
/help	Show help
/retry	Regenerate last response
/undo	Remove last exchange, restore user message to input
/copy	Copy last response to clipboard
/clear	Clear chat history
/personality	List/switch personalities
/config	List/switch config profiles (restarts app)
/impersonate, /generate, /gen	Generate suggested user response (long-form)
/imps	Short impersonate (under 15 words)
/suggest	Toggle auto-suggest after responses
/thinking	Toggle reasoning at inference level
/stream	Toggle streaming mode
/project	Toggle project prompt merge
/prompt	Show current system prompt
/sys <msg>	Inject a system message (alias: /system)
/model	Show current model
/stats	Show generation statistics (TTFT, t/s, tokens)
/compact	Summarize conversation to free context
/context	Show context info

Configuration

Config lives in ~/.config/ochat/:

• config.conf - Default settings (host, model, context size, etc.)
• *.conf - Named config profiles
• personalities/ - System prompt templates (.md files)

Multiple config profiles

You can create and switch between different configurations (different models, personalities, settings):

ochat --new                          # Create a new profile from scratch
ochat --use-config my-creative       # Use profile for this session only
ochat --use-config my-creative --as-default  # Make it the new default

Or switch in-app with /config. When switching with --as-default, your current config is backed up automatically.

Bundled personalities (copied on first run):

• default - Helpful, concise assistant
• creative - Brainstorming and unconventional ideas
• storyteller - Narrative and creative writing

Advanced model options

You can add any Ollama model parameter in config.conf. These are passed directly to the API without validation:

[model_options]
temperature = 0.7
top_p = 0.9
top_k = 40
min_p = 0.05
repeat_penalty = 1.1
# Any other Ollama option works too

See Ollama docs for available options. Not exposed in the setup wizard - edit the config file manually.

Backend modes

Configure in config.conf under [defaults] backend: ollama, openai, llama_cpp, or auto.

Ollama (default): Full feature support — model listing, num_ctx, model_options, real token tracking.

OpenAI-compatible: Works with LM Studio, vLLM, text-generation-inference, and other servers exposing /v1/chat/completions. Uses the openai Python client. Limitations:

• No interactive model selection (model listing is used for connection testing only — you must configure the model name manually)
• num_ctx and model_options are ignored (server-side settings apply)
• Context usage tracking is disabled (percentages and warnings don't apply)
• Setup wizard won't work (configure config.conf manually)

Llama.cpp: Uses llama.cpp server's /v1/chat/completions and /info endpoints. Supports num_ctx via /info and real token tracking via include_usage.

Auto: Tries Ollama → llama.cpp → OpenAI in sequence.

To use with LM Studio:

[defaults]
backend = openai

[server]
host = http://localhost:1234

[defaults]
model = your-loaded-model-name

Self-signed certificates (homelab)

For HTTPS hosts with self-signed certificates, the setup wizard will detect SSL errors and offer to disable verification. You can also set it manually in config.conf:

[server]
host = https://my-homelab:11434
verify_ssl = false

The greeting will show the backend type: Connected (Ollama), Connected (OpenAI), Connected (llama.cpp), or Connected (auto).

Debugging

Logging is disabled by default. Use -d to enable:

ochat -d

Log files are written to temp (auto-cleaned after 7 days):

• Unix/Mac: /tmp/ochat-YYYYMMDD-HHMMSS.log
• Windows: %TEMP%\ochat-YYYYMMDD-HHMMSS.log

Code quality

Average cyclomatic complexity: A (3.3) — no D or F.

Functions rated C (radon cc)

Function	File	CC	Grade
run_setup	config.py	18	C
_handle_undo	commands.py	14	C
main	app.py	14	C
_consume_chunks	generation.py	13	C
_handle_config_command	commands.py	12	C
_handle_command	commands.py	12	C
switch_config_to_default	config.py	12	C
_handle_compact	commands.py	11	C
_status_text	app.py	11	C

Performance notes

Textual is lovely for building TUIs but its default paths don't hold up under realistic LLM streaming workloads — long streamed Markdown plus a hidden Chain-of-Thought block exposes a few pathologies that had to be worked around to keep the UI responsive:

• Streaming Markdown rebuilds: the stock Markdown widget re-parses the entire buffer on every append(), blocking the event loop on long reasoning. Subclassed to run the parse in a thread pool (StreamingMarkdown), so chunks keep flowing while mistune chews.
• :hover on large subtrees: a plain :hover CSS rule on the reasoning block triggers a full stylesheet re-apply to every descendant on every mouse-move (thousands of MarkdownBlock widgets, deep selector matching). Replaced with inline-style toggles in on_enter / on_leave and manual descendant cache invalidation — zero cascade cost, uniform tint.
• Fire-and-forget AwaitComplete: Markdown.update() returns an AwaitComplete that schedules its task via asyncio.gather() inside __init__. Calling it without await leaves an orphan task; at shutdown or lock contention you get CancelledError was never retrieved and can hang. Every call site now awaits.
• Full-replace vs. incremental in non-stream mode: forwarding the accumulated reasoning to Message.update(reasoning=...) on every chunk did an O(chunks × len) Markdown re-parse + remove-then-mount of every block. Reasoning is now rendered once at finalize, and the spinner body update is throttled to 10 Hz (Markdown.update is ~20-50ms of DOM churn per call).
• Visual-style cache on descendants: Textual caches per-widget composited colors in _visual_style_cache. Inline style changes on a parent don't invalidate children's caches, so hover tint would only apply to empty cells, not text cells. Fixed by walking descendants and calling notify_style_update() on hover enter/leave.

Net effect: reasoning streams with reasoning collapsed stay indistinguishable from plain text, hover-on-reasoning no longer freezes the UI for minutes, and non-stream generations render in real time instead of catching up several seconds after completion.

Alternatives

If you want a more full-featured Ollama TUI, check out parllama and oterm. Both are capable and well-maintained. Their busier interfaces weren't to my taste — this project exists because I wanted a clean, single-screen interface with no panels or sidebars to manage.

License

Public domain. You can't copyright AI-generated code anyway.

---

"Vibe coding is awful, LLMs steal real code, you should've coded it yourself"

Frankly my dear, I don't give a damn.