🦞🧪 TensorClaw is an ML research harness on top of pi, inspired by autoresearch.

pip install -e .

Requires:

• Python 3.10+
• pi on your PATH (brew install pi-coding-agent)
• OPENAI_API_KEY or ANTHROPIC_API_KEY

tensorclaw [--dry-run]

Run from the target repo root.

TensorClaw enforces a metric-gated loop:

1. Chat with the agent.
2. Agent proposes one experiment plan.
3. You approve (Y) or reject (N).
4. Harness runs the experiment and parses objective metrics.
5. Harness keeps improvements, reverts regressions/crashes, and records the iteration.

• Chat pane: streamed user/agent conversation + pending proposal card
• Metrics pane: objective sparkline, live monitor, run context, and pi token usage
• Iterations pane: baseline + keep/discard/crash history with best/latest markers
• Output pane: streamed setup/apply/experiment output
• Status bar: phase, metric snapshot, and total pi tokens

TensorClaw stores run state under .tensorclaw/ in the target project:

.tensorclaw/
  results.tsv         # iteration ledger
  journal.md          # readable run journal
  metrics.jsonl       # live metric samples
  memory.jsonl        # retrieved memory for future prompts
  logs/               # run/apply logs
  instructions/       # saved prompts/instructions
  session.jsonl       # chat + pending plan state

We use the TinyStories dataset. See autoresearch for more info.

For a demo on Apple Silicon, run TensorClaw from your autoresearch-mlx project root:

tensorclaw

MIT