This page separates shipped behavior from designs that are not implemented.
Current Limitations
Linear execution only
Every tool runs its steps in fixed order. Chaining and condition fields are stored but ignored. There is no skip, branch, loop, retry, fallback, or resume-from-step behavior.
No tool-to-tool calls inside a tool
One tool cannot invoke another tool as a step. The Automations tab can sequence multiple tools, but the beta automation runtime also flattens them into one sequential list.
Saved references are ignored
Deployments can store Reference entries and the interface displays them. At run time, ThinkForge warns that reference sources are not read, then continues without them.
Put required source material in Input Sources until reference resolution is implemented.
MCP cannot run Agent Tools
ThinkForge MCP does not expose tool listing, inspection, or execution.
MCP handles project knowledge, search, safe document changes, artifacts, tags, indexing, statistics, and reconciliation.
Chrome integration is incomplete
The extension's Tools tab launches the desktop Tool Dashboard rather than presenting tools in Chrome.
Extension code can start a tool, but it is fire-and-forget and does not return the completed result. Listing tools from the extension is not implemented.
Only two trigger types
Schedule and new-file watch triggers are shipped.
There is no application-event trigger. File watch observes created files only, does not include subfolders, and has no debounce or copy-completion wait.
Incomplete execution history
Durable logging is not applied uniformly. Trigger-fired runs are logged; runs from other surfaces are not, so the stored history is not a complete audit trail.
Limited error recovery
A failing step ends the run. There are no automatic retries, failure branches, or partial resume. Cancellation is checked between steps. Most active skills finish before cancellation takes effect; webpage retrieval cancels active network work, and a Python step terminates its process tree.
Settings are often validated at run time
Some service targets and credentials are collected on first use. There is no comprehensive preflight check before a desktop run.
Python execution is deliberately bounded
RunPythonScript runs with your current Windows account permissions and is not sandboxed. ThinkForge does not install Python or packages, discover an interpreter, manage environments, inspect scripts, or repair failures. Trust covers the selected script file, not imported modules or dependencies. A tool can contain only one Python step.
MCP-backed AI is local and bounded
ProcessWithThinkForgeMCP can call selected operations from ThinkForge's local MCP catalog, but it is not a general MCP client. Every step needs a non-empty allow-list and can perform at most ten model-and-operation rounds. The configured model route must support tool calling, and failures do not fall back to an ordinary AI response.
Headless web retrieval and browser limits
FetchWebPage retrieves public static HTML, XHTML, or plain text. It does not execute JavaScript, use a signed-in browser session, access private networks, or bypass login, CAPTCHA, paywall, or bot-protection flows. It supports up to 25 supplied addresses but does not crawl follow-on pages.
Interactive browser automation skills are not available to the tool planner. Browser page capture is handled by the Chrome extension and Power Strip instead.
Workflow edits are live
Changing the reusable tool's step sequence changes future runs for every deployment and automation that uses it. There is no version history, no pinning a deployment to a specific revision, and no rollback.
This is different from Tool settings on a selected deployment. Deployment settings change only that placement. Settings opened on a reusable tool supply defaults for deployments created later.
Designed but Not Shipped
The following are future work, not beta features:
- General Agent Tool discovery and execution through MCP
- Per-tool MCP visibility
- Structured MCP execution input and result polling
- Complete execution history through MCP
- Conditional tool steps
- Tool-to-tool composition
- Browser automation skills
- Reference-source resolution
- Definition versioning and rollback
- Comprehensive readiness checks and retry policies
Related Automation Limitations
Automation connections currently determine run order only. Conditions, true branches, fan-out, merge nodes, and automation triggers are planned. See the Automations section for current behavior.