An open-source
studio at the
boundary of
LLMs & software.

Prompt engineering without version control is archaeology. Pressmark treats prompts as first-class artifacts: every commit gets a content-addressed SHA, a parent pointer, and a commit message.

Five built-in scorers (exact, contains, regex, LLM-judge, semantic). Async eval runner with crash-safe per-row writes. Compare any two runs side-by-side. CI gate via --min-pass-rate.

The fastest path between a terminal and any frontier model. Streams markdown as it arrives. Switch models mid-conversation. Tracks tokens and USD per message.

Built on Textual. SSE streaming via httpx. Searchable model browser, system-prompt editor, export-to-file. One pip install, no config.

Speech in, speech out. Microphone, transcription, model, synthesis, speaker — wired into one streaming loop. Defaults to local-only; cloud providers are opt-in per stage.

Sentence-boundary buffering means TTS begins speaking the first sentence while the LLM is still generating the second. Every stage is swappable with a single config line.

The open-source meeting assistant. Cluely-class stealth overlay invisible to screen-share, agent-driven answers, real-time transcript pipeline, scriptable HTTP daemon under everything.

Bring your own model — Ollama or OpenRouter, side by side. Ten built-in scenario templates: coding, meeting, interview, sales, email, writing, research, math, custom, plus a smart-detect Auto. Download as a desktop app, install the Python CLI, or hit the daemon's REST endpoints from anywhere.