Does it make sense to ship an MCP code mode API? I'm surprised you're recommending MCP as-is when concerned about context usage optimization. I don't have a lot of hands-on experience either way yet so I'm curious what's best and/or most popular... I understand MCP is less effort and still affordable at VC-subsidised prices.
for the integration piece that ties into Claude Code and other places where AI is used most frequently? yes I think it does... we're not fighting context in Opus/Haiku as much as we are in smaller models. Smaller models have a more direct/tight interface that doesn't bloat the tool space in my experimentation
Interesting, I built a ticketing system similar to Beads which has yielded more predictable results with Claude and other models, and I'm currently building a custom harness, I'm able to use offline models though my GPU ram bandwidth is much lower, but I'm also planning on doing something similar to what you've built, namely the editing tools and what not, I hate how long it takes for Claude to look for files, it feels wasteful. I'm still astounded that everyone else has figured out ways to speed up harnesses, but Claude Code is still slow like a slug. I don't even care if I am waiting on the LLM in terms of slowness, but running local tools slowly bothers the living crap out of me, stop using grep, RIPGREP IS FASTER!
In any case, I'll have to check out Statewright after work ;)
Does it make sense to ship an MCP code mode API? I'm surprised you're recommending MCP as-is when concerned about context usage optimization. I don't have a lot of hands-on experience either way yet so I'm curious what's best and/or most popular... I understand MCP is less effort and still affordable at VC-subsidised prices.
for the integration piece that ties into Claude Code and other places where AI is used most frequently? yes I think it does... we're not fighting context in Opus/Haiku as much as we are in smaller models. Smaller models have a more direct/tight interface that doesn't bloat the tool space in my experimentation
Interesting, I built a ticketing system similar to Beads which has yielded more predictable results with Claude and other models, and I'm currently building a custom harness, I'm able to use offline models though my GPU ram bandwidth is much lower, but I'm also planning on doing something similar to what you've built, namely the editing tools and what not, I hate how long it takes for Claude to look for files, it feels wasteful. I'm still astounded that everyone else has figured out ways to speed up harnesses, but Claude Code is still slow like a slug. I don't even care if I am waiting on the LLM in terms of slowness, but running local tools slowly bothers the living crap out of me, stop using grep, RIPGREP IS FASTER!
In any case, I'll have to check out Statewright after work ;)
Pretty cool. Looks like stately.ai but catered towards agentic state machine workflows. Really interesting!