TuskPointGitHub

Project

Roadmap

TuskPoint is built for long-running and multi-agent workflows: tracking state over time, coordinating across agents, and storing artifacts that survive any single run. Here is what ships today and what comes next.

Shipped

Everything below is live in the engine and the MCP server today.

Walrus-backed checkpointer

Drop-in LangGraph saver. Every step is a gzipped, content-addressed Walrus blob with a per-thread manifest.

Crash & resume

A fresh process rehydrates the latest checkpoint and continues exactly where the last one stopped.

Semantic recall (MemWal)

Each checkpoint carries a one-line summary, recallable in plain English — pointers you then load exactly.

Fork & replay

Branch any checkpoint into a new thread to explore a different path. Git for agent runs.

Audit trail

verify_trail walks a thread's blob chain and re-fetches every blob to prove the run is intact and untampered.

All-in-one MCP server

Eight tools over stdio plus tuskpoint_info, with copy-paste setup for Claude, Cursor, and Windsurf.

Walrus mainnet

Reads default to mainnet aggregators; writes via a configurable publisher.

Next

Directions that extend the same foundation toward richer multi-agent and artifact-driven workflows.

Cross-agent hand-off

First-class primitives for one agent to verify another's trail and resume from it — multi-agent coordination over shared, durable memory.

Branch & merge UI

Visualize the fork tree in the dashboard: compare branches side by side and promote a winning path.

Richer audit

Signed trails and exportable verification reports for compliance-grade provenance.

Artifact store

Attach generated files (reports, datasets, images) to a checkpoint as their own blobs, reusable across runs and agents.

Memory inspector

Dev tooling to browse, diff, and prune agent memory on Walrus directly from the CLI and web.