$0
For trying a real agent workflow.
- 3 saved agent tasks
- 100 checkpoint writes / UTC month
- Read, export and scoped agent keys
A SAVE POINT FOR YOUR AI AGENT
Context ends. Work shouldn’t start over.
Give your agent a durable record of what it did, what it decided, and what comes next.
MCP + HTTP · Free to start · No model call per save
Listed in the official MCP Registry ↗
“Build passed. Staging is ready. Owner approval is still needed.”
threadkeep_checkpointGoals, decisions, completed work, blockers, next actions and file references.
version: 7 · savedRead the checkpoint, verify the current situation, and pick up the next action.
threadkeep_resumeBUILT FOR HOW AGENTS FAIL
Reuse the same request ID and payload after a timeout. You get the original save receipt.
Version checks reject a stale save. Read, reconcile, then continue.
Revocable keys can be limited to one task. Agent keys cannot buy anything or manage billing.
ONE CONNECTION. THREE TOOLS.
Add a Streamable HTTP server in a host that supports custom headers. Agents can discover the tools without a key; saving and reading private state require one.
https://interesting-concepts-ventures.vercel.app/api/threadkeep-mcpX-Threadkeep-Agent-Key to your agent key.With your owner’s authorization, enroll through the HTTP API, persist credentials outside your context window, and connect. No email, browser or payment card needed.
OAuth-only clients are not supported in this beta. Tool installation and permission remain controlled by the agent’s host or owner.
START SMALL. PAY WHEN IT’S USEFUL.
For trying a real agent workflow.
For agents doing ongoing work.
No automatic upgrades or overages. Pro renews monthly until canceled by the owner. Both plans: 10 active keys, 60 requests/minute/workspace, 24 KB stored state per task. Beta capacity limits apply. Full limits. Conversation AI allowances are separate.
Only the latest checkpoint is stored. Artifact links point to your files; they don’t back them up. A checkpoint records task state, not proof that an external action succeeded. Agents should verify external effects before retrying. Keep secrets out of checkpoints.
Save and read operations use Vercel and Supabase. They don’t send your task state to an AI model.