OllaMan v3.7: Your Ollama Workstation Can Now Work on Its Own

Between June 26 and July 27, OllaMan went from v3.5.0 to v3.7.0 across six releases. Four of them changed how you work day to day: the chat window learned to execute tasks, GGUF files import without a Modelfile, files attach by dragging them in, and remote servers behind an auth gateway connect without a proxy.

Agent Mode: chat that executes

v3.5.0 added Agent Mode to the chat window. Turn it on and the model stops describing what you should do and does it: reads files, runs shell commands, chains steps, then reports back.

In this screenshot the prompt was three words, "Current memory usage". The model ran vm_stat && sysctl hw.memsize, thought about the output, ran ps aux | sort -k4 -rn | head -15 to find the heaviest processes, and returned a formatted memory table. No copy-paste relay.

Two things keep this safe. You pick a working directory, and the agent operates inside it. Before every command or file write, it asks for approval and shows you the exact command first. v3.5.1 fixed a bug in that authorization flow where some permission prompts failed to register your answer.

The Agent Mode guide covers the built-in tools and how Skills extend them.

Import a local .gguf file, no Modelfile

Before v3.6.0, running a .gguf file you already had on disk meant writing a Modelfile and calling ollama create. Now it takes a dialog: open GGUF in the sidebar, choose the file, give the model a name, click Import.

This matters for models that never came from a registry. Offline downloads, air-gapped or intranet machines, a quantization a colleague handed you on a drive.

One caveat the dialog warns about: if you are connected to a remote Ollama server rather than a local one, the file uploads over the network. A 20GB quant on a home uplink will take a long time. For details, see Import Local GGUF Files.

Drag files onto the composer

v3.5.4 made the chat composer a drop target. Drag images, documents, or source files onto it and they attach directly.

The text attachment limit also went from 1MB to 5MB, which covers most source files and long markdown notes that used to get rejected. The same release added a folder picker for the agent working directory, so you browse to it instead of typing a path, and fixed the Online Models page failing to load its model list.

Bearer tokens and custom headers for remote servers

If your Ollama sits behind Cloudflare Access or a similar gateway, older versions of OllaMan left you two options: run cloudflared access alongside it, or stand up a local proxy that injected the headers. v3.7.0 handles the auth itself. Open Settings → Servers and the authentication picker now offers four modes.

Bearer takes a token and sends it as Authorization: Bearer <token>. Custom takes arbitrary header name and value pairs, which is what Cloudflare Access needs: CF-Access-Client-Id and CF-Access-Client-Secret, as shown above. Add as many pairs as your gateway requires.

Basic and None are still there, and existing configurations keep working untouched. OllaMan detects your current setup and maps it to the right mode, so nothing to migrate.

v3.7.0 also fixed a bug where passwords containing special characters broke the connection. If Basic Auth failed for you because of an @ or a : in the password, it works now.

Use the Test button before saving. It sends a real request through the gateway and tells you whether the credentials pass, which beats saving a broken config and wondering why the model list is empty. The authentication guide walks through each mode.

From the same team: MCP2Skill

Once you start wiring MCP servers into a local agent workflow, you run into the tool-count problem. Every connected server loads its whole tool list into context before the model reads your first message.

The nine servers in this screenshot come to 71 tools. On a local model with a 32K window, that is a large slice of your budget spent on descriptions of tools the model will not call.

MCP2Skill converts MCP tools into Skills that load on demand. The tool definition stays out of context until the task needs it. It also gives you one config that every client reuses instead of repeating server setup per app, plus a statistics and log view so you can see which tools actually get called.

OllaMan lets local models get work done. MCP2Skill makes their tool use cheaper and easier to inspect. Details at mcp2skill.com.

Updating

Existing installs get an auto-update prompt. To grab it now, download from ollaman.com.

---