ChatMol Lab 1.0.0

ChatMol Lab v1.0.0

First public release of the desktop workbench, and the first release built from the open-source repository. macOS (Apple silicon) and Windows (x64) installers are ad-hoc signed but not notarized; on macOS right-click → Open the first time.

Highlights

• Bring your own model. DeepSeek, OpenAI, Anthropic, Google Gemini, OpenRouter, or any OpenAI-compatible server (Ollama, vLLM, LM Studio). Keys are stored per provider; a cheaper "fast model" handles titles, reports, and the tool reviewer.
• No account needed, nothing to pay us. The desktop app opens straight into a local workspace, with no sign-in. Compute runs on the credentials you configure, billed by those providers directly. Connecting ChatMol Cloud is optional.
• No GPU needed. OpenFold2/3, Boltz-2, RFdiffusion, ProteinMPNN, DiffDock, ColabFold MSA, GenMol, MolMIM and Evo2 through NVIDIA NIM; industrial antibody and ADMET workflows through WeMol; background GPU jobs through the ChatMol Bio API with your own key.
• Structure analysis built in. analyze_structure answers the routine questions locally through biotite and Biopython: secondary structure, superposition and RMSD with sequence-matched chains, interface contacts with buried area and a clash warning, per-residue solvent accessibility, pLDDT or B-factor confidence, and sequence properties. The two Python packages install themselves into the bundled runtime the first time it runs.
• Molecular viewers. Select residues in the built-in Mol* viewer and attach them to the conversation. Drive a live PyMOL or ChimeraX window from chat through bundled MCP servers (Settings → General → Molecular viewers).
• Project folders. Point a session at any folder on disk; files, scripts, and viewer images land there.
• Auto tool review. Read-only shell commands run immediately, destructive ones ask, everything else is checked by the fast model against your task and the workspace boundary. Manual, reviewer, and unrestricted modes remain available. Where the operating system offers no sandbox, every shell command asks first.
• Windows. Native conda runtime plus an offline WSL2 Linux runtime for bioconda tools.
• MCP servers. Any stdio MCP server can be added; its tools appear to the agent as mcp__<server>__<tool>.

Removed

• ESMFold. NVIDIA withdrew the NIM endpoint; it answers 404. Use OpenFold2 for a single sequence or Boltz-2 for complexes.

Known limitations

• PyMOL / ChimeraX servers need the mcp Python package in the bundled runtime (one-click install in Settings).
• Installers are ad-hoc signed, not notarized. macOS: right-click → Open on first launch; on macOS 15 and later, open it once and then allow it under System Settings → Privacy & Security. Windows: SmartScreen → More info → Run anyway.
• Full 8-state DSSP secondary structure needs mkdssp (Settings → Structure analysis → Add full DSSP); without it the assignment is 3-state.
• Session sync to the hosted server is optional and off by default.

Also released

• chatmol-mcp-servers (MIT): the PyMOL, ChimeraX, database, and NVIDIA NIM servers as a standalone package for DeepSeek Harness, Claude Code, Codex, and Cursor.