2nd Place
ToolForge
ToolForge is an autonomous AI agent that solves a core limitation of traditional AI: being stuck with only the tools it was given.
Core trick
Planner checks a cached tool registry first, otherwise a builder agent writes and sandbox-tests a new tool, then stores it.
Description
ToolForge is an autonomous AI agent that solves a core limitation of traditional AI: being stuck with only the tools it was given. Instead of a fixed set of integrations, ToolForge starts with nothing and builds what it needs on the fly. Its Planner agent checks existing tools first, then either routes through Composio's 150+ integrations or spins up the Builder agent to write custom code from scratch, test it, and verify it — no human needed.
Every tool built gets stored in a Snowflake registry, so the agent compounds knowledge over time. Orchestrated by CrewAI, reasoned over by Llama via Snowflake Cortex, and capable of autonomously paying for APIs through Skyfire — ToolForge is one of the few agents that truly operates end-to-end.
README excerpt
# ToolForge
A self-extending agent that builds, tests, and caches its own tools.
Launch the terminal UI with `python -m toolforge.tui`.
## What it does
ToolForge is an agent that **builds its own tools on demand and remembers them**.
When a request comes in, ToolForge first checks whether it already has a tool for the
job. If it does, it runs it. If it does not, it builds one, tests it in a sandbox, saves
it, and runs it. Every request that misses the cache makes the system permanently
smarter.
A ToolForge "tool" is one of two things:
- **A Composio plan** — an ordered list of Composio tool calls. Anything that touches an
external service (email, GitHub, calendars, Slack) is reached through Composio, which
handles all authentication. ToolForge never hand-writes API code or asks for a
credential.
- **A code tool** — a small piece of pure-logic Python that touches no external service:
maths, parsing, format conversion, date arithmetic.
The registry of every tool ever built lives in Snowflake, which also powers the
"do I already have this?" semantic search.
## Architecture
Four moving parts: **CrewAI** orchestrates, **Composio** connects to the outside world,
**Snowflake** is the long-term memory, and an **LLM** does the thinking.
```mermaid
graph TD
USER["User request"] --> DECOMP["Decomposer<br/>split into ordered sub-tasks"]
DECOMP --> PLANNER["Planner<br/>route each sub-task"]
PLANNER --> REG[("Snowf