Forge

ToolSmith is a procedural memory layer for AI agents that turns repeated work into tested, reusable tools. It watches how people and agents work, both on screen through a Chrome extension and a frame interpreter and in structured logs, and spots the workflows they repeat week after week. It only proposes a tool when a workflow truly recurs, and it visibly declines one-off binges and inconsistent routines. For each accepted pattern it writes the specification, code, tests and a plain-language tutorial. The tool must reproduce the user's real past results in a network-isolated sandbox before anyone can approve it. Once in use, a tool earns autonomy step by step, from dry-run to supervised to autonomous, and is demoted instantly on failure. When a site's layout changes and a tool breaks, ToolSmith detects the drift, diagnoses the change, and releases a fixed version that passes on both the old and new page. It also tunes its own guardrails: stricter settings apply automatically, while looser ones wait for the user's approval. MongoDB Atlas runs the whole system: - Hybrid vector search ($rankFusion) and multimodal embeddings recall past work and match screenshots. - $graphLookup resolves tools built from other tools. - Change streams drive the background workers. - Multi-document transactions make promoting a tool all-or-nothing. The result: a repeated nine-minute, 18,000-token task becomes a verified tool that runs in seconds.