Kanon

Your LangGraph agent crashed at step 38. Kanon resumes it at step 38.

# Before
from langgraph.checkpoint.memory import MemorySaver
graph = builder.compile(checkpointer=MemorySaver())

# After — agents resume from last signed checkpoint on restart
from kanon.langgraph import KanonCheckpointer
graph = builder.compile(checkpointer=KanonCheckpointer(api_key="kn_..."))

pip install kanon[langgraph]

Every checkpoint is cryptographically signed. If the stored state is altered between save and resume, Kanon raises KanonTamperedError before returning state to LangGraph.

Run the demo →

Crash-resume demo

$ python demo/crash_resume_demo.py --demo

[kanon-demo] Starting 5-step research agent  thread=demo-thread-1
  Step 1/5  node=gather_sources        ✓  snap=snap_a1b2c3  (signed)
  Step 2/5  node=summarise_sources     ✓  snap=snap_d4e5f6  (signed)
  Step 3/5  node=draft_findings        ✓  snap=snap_g7h8i9  (signed)
💥 Simulated crash after step 3. Process state wiped.

[kanon-demo] Resuming  thread=demo-thread-1  from=snap_g7h8i9
  Step 4/5  node=verify_findings       ✓  snap=snap_j0k1l2  (signed)
  Step 5/5  node=write_report          ✓  snap=snap_m3n4o5  (signed)

[kanon-demo] Complete. Signed receipt chain:
  snap_a1b2c3  →  snap_d4e5f6  →  snap_g7h8i9  (crash)  →  snap_j0k1l2  →  snap_m3n4o5
  proof_bundle hash: e3b0c442...  agent=kanon-demo  tenant=demo

Or as two separate processes:

bash demo/run_demo.sh

How it works

KanonCheckpointer is a drop-in replacement for LangGraph's MemorySaver. After each node, LangGraph calls put() — Kanon signs the checkpoint and stores it server-side. On resume, get_tuple() calls restore_snapshot, the server verifies the HMAC proof bundle, and LangGraph continues from the last good state.

The signed receipt chain links every checkpoint to its predecessor. After a crash, the resume linkage (work_artifacts[resume_linkage]) keeps the server-side audit graph connected across the recovery boundary.