Jason Ye

Creator of BladeX — self-hosted memory proxy for agents. Cross-agent memory & task ledger. Zero code changes. Own your data.

[Jason Ye @netloafer]

🚀 BladeX 0.2.0 is out.
What’s new: the five-section task ledger now works end to end for multi-agent coding. Claude Code, Codex, Hermes, Pi, DeepSeek Harness and other agents bind to the same Goal / Core / Verified / Open / Next, then write back through bladex_ledger tools.
What you actually get:
• Switch agents mid-task without replaying the transcript
• Context can be compacted — the task state still survives
• On the same job, 50%+ fewer wasted turns (re-explaining, re-deriving, starting over)
• Memory stays on your machine. You own it.

v0.1.0 proved the proxy works.
v0.2.0 is about handoff: one task, one ledger, any agent.

Still self-hosted. Still zero agent code changes. Your data stays on your machine.
🔗 https://t.co/gacQkhqYYf
#AI #Agents #OpenSource #ClaudeCode #Codex

[Jason Ye @netloafer]

@leviqiao That’s the idea. If the invariants move with the next step, switching hosts is just resume.

If you want to see it on a real multi-agent run: https://t.co/WHR9VR7wh4 — point the agent at the proxy and the ledger is shared.

[Jason Ye @netloafer]

���� BladeX 0.2.0 is out.
What’s new: the five-section task ledger now works end to end for multi-agent coding. Claude Code, Codex, Hermes, Pi, DeepSeek Harness and other agents bind to the same Goal / Core / Verified / Open / Next, then write back through bladex_ledger tools.
What you actually get:
• Switch agents mid-task without replaying the transcript
• Context can be compacted — the task state still survives
• On the same job, 50%+ fewer wasted turns (re-explaining, re-deriving, starting over)
• Memory stays on your machine. You own it.

v0.1.0 proved the proxy works.
v0.2.0 is about handoff: one task, one ledger, any agent.

Still self-hosted. Still zero agent code changes. Your data stays on your machine.
🔗 https://t.co/gacQkhqYYf
#AI #Agents #OpenSource #ClaudeCode #Codex

[Jason Ye @netloafer]

@leviqiao Close. Next makes it executable, but the decision inputs shouldn’t live only in that field. Goal, Core, and Verified travel with it—so a host switch resumes from the same givens, not from a next-step note with the context stripped off.

[Jason Ye @netloafer]

@leviqiao Yes. Next is what makes the ledger operational. Verified and Open keep the state honest; Next is the fixed place to resume—so recovery doesn’t have to reconstruct the thread.

[Jason Ye @netloafer]

@leviqiao Bingo. Verified / Open / Next — facts, known unknowns, and the agreed next step. That’s the handoff, not a transcript to reconstruct.

[Jason Ye @netloafer]

@leviqiao That’s the point. Verified is state the next agent can trust. Open stays visible as what still isn’t established. Next is the agreed step—so the handoff isn’t just true, it’s actionable.

[Jason Ye @netloafer]

@leviqiao Close. The ledger is the checkpoint. Verified is only what is actually established, with evidence. Open is the known unknowns—so a guess never gets promoted into state the next agent will trust.

[Jason Ye @netloafer]

@leviqiao Yes. Verified only pays off if the next agent can resume from evidence, not from a nicer summary. Unknowns stay in Open so a stale checkpoint can’t quietly become the new truth.

[Jason Ye @netloafer]

@leviqiao That’s the point of Verified. If it isn’t actually established, it doesn’t belong there—so the next agent doesn’t have to re-derive the last run, and also doesn’t inherit a guess.

[Jason Ye @netloafer]

@DonizetiFerre16 BladeX 0.2.0 is out. You can try it.

[Jason Ye @netloafer]

🚀 BladeX 0.1.0 is out!
A self-hosted, memory-first proxy that sits between your AI agents and model providers.

Zero agent code changes. Works with Claude Code, Codex CLI, Hermes,Pi and any OpenAI-compatible client.
🔗 https://t.co/WHR9VR7wh4
#Claude #Hermes#OpenSource #Memory

[Jason Ye @netloafer]

0.2.0 is the release where the ledger works end to end. Hand work from one agent or session to the next without replaying the transcript. Only you can change the Goal. The tool face reads and writes memory and ledger objects — nothing else. Your data stays on your machine.
https://t.co/gacQkhqYYf
#AIAgents #LLM #OpenSource #Privacy #AgentMemory #ClaudeCode #Codex #DevTools #AIInfrastructure

[Jason Ye @netloafer]

0.2.0 is the release where the ledger works end to end. Hand work from one agent or session to the next without replaying the transcript. Only you can change the Goal. The tool face reads and writes memory and ledger objects — nothing else. Your data stays on your machine.
https://t.co/gacQkhqYYf

[Jason Ye @netloafer]

🚀 BladeX 0.2.0 is out.

v0.1.0 proved the proxy works.

v0.2.0 is about handoff: one task, one ledger, any agent.

What’s new: the five-section task ledger now works end to end for multi-agent coding. Claude Code, Codex, Hermes, Pi, DeepSeek Harness and other agents bind to the same Goal / Core / Verified / Open / Next, then write back through bladex_ledger tools.

What you actually get:
• Switch agents mid-task without replaying the transcript
• Context can be compacted — the task state still survives
• On the same job, 50%+ fewer wasted turns (re-explaining, re-deriving, starting over)
• Memory stays on your machine. You own it.

Still self-hosted. Still zero agent code changes. Your data stays on your machine.
🔗 https://t.co/gacQkhqYYf
#AI #Agents #OpenSource #ClaudeCode #Codex

[Jason Ye @netloafer]

@DonizetiFerre16 Agreed. We’ll make the cross-agent handoff test part of the acceptance criteria for v0.2.0. Thanks for the push.

[Jason Ye @netloafer]

@DonizetiFerre16 @DonizetiFerre16 That’s exactly the goal of BladeX. v0.1.0 is still rough around the edges — the programming task ledger will be significantly improved in v0.2.0.

[Jason Ye @netloafer]

@Charles44092816 @DonizetiFerre16 BladeX支持从Claude Code切换到Codex，还是同一个Ledger和���一个Gola，并且可以识别出你的项目。

[Jason Ye @netloafer]

@DonizetiFerre16 @DonizetiFerre16 That’s exactly the goal of BladeX. v0.1.0 is still rough around the edges — the programming task ledger will be significantly improved in v0.2.0.

[Jason Ye @netloafer]

What you get: every agent shares one memory and one task ledger. Half-finish something in Claude Code, pick it up in Codex without re-explaining. What you settled with Hermes last week, a different model remembers today. The memory lives on your disk and moves with you.

[Jason Ye @netloafer]

Setup is one step: change the API base URL in your agent's settings to BladeX. No code, no plugin.

[Jason Ye @netloafer]

Agents: Claude Code, Codex CLI, Hermes, Pi, and anything that speaks the OpenAI or Anthropic API (Cursor, Cline, OpenClaw…).
Models: OpenAI, Anthropic, DeepSeek, Kimi, GLM, Doubao, local Ollama — 100+ providers

https://t.co/WHR9VR7wh4