DataClaw

DataClaw is a data-science harness on a swappable agent runtime.¹² Its domain surface holds ledgers, hypotheses, findings, and notebooks.²

Overview#

Governance owns history, tools, hooks, and approvals, and the runtime owns the loop.² A run bundle is frozen for the turn, and claims bind to evidence before publication.²

Mechanism#

1. The analytics surface sits on a swappable runtime.²
2. Tool calls are rechecked against project scope, then against hooks and approvals.²
3. The run bundle is frozen so that a reload cannot mix providers mid-turn.²
4. Claims are bound to evidence and reviewed, and only bounded repair is allowed.²
5. Methodology and ship decisions stay with the human.²

Applications#

The harness is relevant when analytics work needs its own surface without a second architecture.

Limitations#

Local arbitrary code execution and plaintext keys form a trust boundary.²

See also#

• Hermes, GBrain, Autoresearch loop, Verification and stop conditions – pieces it composes with
• Harness runtime – runtime internals
• ClawSweeper – related fleet operations

References#

Footnotes#

1. https://github.com/gopivikranth28/DATACLAW-AGENT ↩

2. https://x.com/gopivikranth/status/2049143600207081647 ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰