Open Source Multi-Harness Orchestrator : Combine and supervise any harness
- Tweets 244
- Following 174
- Followers 16
- Likes 404
Is it ILLEGAL to have Claude Desktop/Codex delegating to opencode x BigPickle ?
You can now have your fav desktop app (anything supporting MCP Apps, thanks ) orchestrating any harness (Claude, Codex, Hermès, Pi, Grok and 12+).
I'm working on a cross-harness framework, and made a pluggable "Agent Apps" system communicating with the daemon with MCP or HTTP.
It allows to create easily an UI that runs on your local harnesses, natively compatible with MCP Apps implementation.
1
2
71
Thxs so much @idosal1 @mxstbr @liadyosef for making MCP Apps so fun to work with ! The potential of this is untapped.
I've burn some tokens to have a plug & play "MCP Host" package, that can take any classic app, MCP-ready.
github.com/agentproto/ts/tre…
3
76
Got one feedback / improvement suggestion @thsottiaux : the GUI is eating lots of my precious memory. Can we reduce it to < 1.5 GB by End of Month ?
10
We had our first external contribution and reached 10 stars this week !
A small step for ego, but a big step for making orchestration cross-harness.
Open-source for the win !
8
Is your agent waiting for your "Check PR status", "Build has failed", "Merge conflicts, fix" ?
Mine is not ! From MCP Events to Claude Channel and Agentproto custom Sentinel, you'll go from static to hectic.
agentproto.sh/blog/get-your-…
2
81
There are 3 reasons why you are burning more tokens than you should : switching models, compacting during a task, and NOT compacting after your lunch.
See our report :
agentproto.sh/blog/why-your-…
11
Your AI agent costs more with every message. We checked 160 real sessions to see how to cut the bill.
Each answer re-reads the whole conversation. Message 40 carries messages 1 to 39.
The most efficient way to slash costs ? Prompt Caching.
The 3 ways to improve it :
3
22
1. Keep the top of your chat stable (instructions, tools, files). Any change there breaks the cache.
2. Stick to one model per session. The cache doesn't carry over.
3. Don't leave a session idle for long. The cache expires after a few minutes.
1
1
16
x.com/claudeai/status/210789… interesting followup
One more thing: we’re halving the price of cache reads on Claude Sonnet 5.5, to $0.10 per million tokens. That makes Sonnet 5.5 around 20% cheaper to run on most long-running work.
5
———
We measured all this on real sessions. Full write-up, method and script:
agentproto.sh/blog/why-agent…
5
