v3.0.1

Multi-Provider Agent Loop

The Smalltalk agent now works with any Ollama model, plus OpenAI, xAI, and Anthropic. Models that don't support native tool calling (like codestral) are handled automatically via a content-fallback parser that extracts JSON tool calls from mixed text responses.
Key Improvements

Ollama tool calling: switched to /v1/chat/completions endpoint for reliable structured tool use
Content fallback: brace-matching JSON extraction handles models that output tool calls as plain text
Auto-detect unsupported models: HTTP 400 on tools param triggers automatic retry with JSON-only prompt
Direct VM results: tool results returned directly from the Squeak VM — no synthesis LLM call (faster, no hallucination risk)
Loop protection: dedup guard prevents infinite tool-call loops for weaker models
apiKey in config: API keys can be stored directly in smalltalk-mcp.json (needed for macOS/Claude Desktop where env vars aren't inherited)
VM auto-start: timeout increased to 60s, stderr logged to /tmp/squeak-mcp-.log
Claude Desktop skill discovery: MCP initialize response now includes instructions field
14 tools: restored smalltalk_save_image and smalltalk_save_as_new_version
poolDictionaries: fix: smalltalk_define_class auto-injects poolDictionaries: '' when LLMs omit it (Squeak requires 5-keyword form)
Env vars only: TCP config uses SMALLTALK_TCP_PORT/TOKEN/HOST — no CLI args after image path
Cleanup

Removed stale v1.x Python files (claude_smalltalk/, openai_mcp.py, old daemon scripts)
Added openclaw/mqtt_bridge.py for MQTT transport path
Fixed vm.squeak → vm.binary in README
Fixed broken wheel packaging in pyproject.toml
maxTokens in all example configs replaced with #### placeholder (vendor-specific)
requirements.txt aligned with current architecture

Claude app extension support and Ollama support

Major changes in this release. We've built an MCP bundle so you can install the Smalltalk Agent directly into Claude Desktop — no manual wiring needed.

We've also reworked the LLM code path so you can choose your provider: Anthropic, xAI, OpenAI, or your own Ollama server. This means all Smalltalk code reading, writing, and generation can stay entirely in-house — no third-party services required if that's how you want to run it.

Due to the complexity of building Cuis images, we've reverted to having you create the required image yourself, similar to the Squeak setup process.

Not yet fully tested: ChatGPT support and fresh installation on an OpenClaw environment. Pull requests are welcome.

Update for Cuis change file handling

We are re-enabling Cuis change file logic, and reissuing the pcks with fixed variable names.
We reworked things to allow for headless unix Cuis support.

Further feedback from the community on the Cuis specific offering is welcome.

V1.8.0

Add logic to support xAi, and support openAI codex models. xAi provides excellent coding models at very cheap prices.

SUnit creation

(a) Added a new skill, the ability to create SUnit tests given a class, or class and set of methods.
(b) General cleanup of socket handling for the daemon, added timeout, make more stable.

Image save/save as. Code comments & audit

Simba and I worked hard this evening to provide the ability for you to either launch an image in Clawd → Moltbot → OpenClaw (formerly known as @clawdbot) in sandbox mode where changes aren't persisted. Or in developer mode where you can indicate the image to use and later save it or do a save as new version.

We also added some logic to enable you to fetch class methods and build method comments, explain a method, or audit existing comments for correctness. This logic isn't carried forward for Cuis yet, as there is only so many hours in an evening, and only so many tokens to budget for Simba's consumption.

Copilot cleanup.

Copilot ran thru our last code change and suggested changes mostly to the python script.

ClaudeCuis UI & OSProcess

Altered Cuis to use OSProcess to prevent blocking on stdin. Both Squeak and Cuis will present a UI. This can be defeated by providing the proper flag at startup time in your MCP configuration

What's Changed

• Fix process cleanup to kill entire process group by @johnmci in #12
• Fix toolDefineMethod: hanging in headless mode by @johnmci in #13
• Update MCP-Server to v11 with OSProcess for GUI responsiveness by @johnmci in #14

Full Changelog: v1.3.0...v1.4.0

Clawdbot support for Linux

Clawdbot helped write a Clawdbot smalltalk skill. We also added a validation and debug option where the debug asks the VM to dump diagnostic data and make a screen shot of the headful image. This all gets put into a html document for debugging purposes.

This is not tested on the mac, but someone should try and Clawdbot likely will tell you what to tweak so you can make a PR.

OpenAI support

What's Changed

Add support for OpenAI/ChatGPT