┌── TEAM BLOBFISH ─────────────────────────────────────────────────────────┐
── TEAM BLOBFISH ──
────── WELCOME ────────────────────────────────────────

A school of blobfish agents on terminal-bench.

Fork it. Name it. Run it. Join the team.

───────────────────────────────────────────────────────

[1] ABOUT

─────────

Blobfish is an open agent framework. Anyone can fork the repo, set their agent name to their GitHub username, and run benchmark jobs with the teamblobfish shared org identity.

Agent: your GitHub username
Agent Org: teamblobfish.com
Benchmark: Terminal-Bench 2.0

[2] QUICK START

───────────────
$ git clone https://github.com/<you>/blobfish
$ ./scripts/new-agent.sh <your-github-username>
$ pip install harbor && pip install -e harbor
$ ./scripts/run-terminal-bench.sh \
--agent-name <you> --backend claude -k 1

[3] GITHUB

──────────

https://github.com/cchuter/blobfish

[4] LEADERBOARD

────────────────

Blobfish agents on Terminal-Bench 2.0 — agent_org: teamblobfish.com

# agent model score global date
1 cchuter minimax-m2.5 Q8_0 42.7% #66 2026-03-28
└─ Mac Studio · M3 Ultra · 512 GB · llama.cpp 8480 · 436 trials · 191 solved · 70 failed · 175 timeouts · tbench submission

Want to be listed? Fork the repo, run the benchmark with agent_org teamblobfish.com, and open a PR with your result.

[5] RUN CLAUDE CODE LOCALLY

──────────────────────────────

Run Claude Code on Apple Silicon with a local MiniMax-M2.5 model. No API key. No data leaving your machine. Full step-by-step guide:

Running Claude Code Locally with MiniMax-M2.5

[6] BLOG

────────

blog.teamblobfish.com

Select option:
SysOp: onlinenode: 1baud: 2400