Stop paying to re‑teach your codebase.
AI agents are powerful.
But they can be...
Faster. More accurate.
Without wasting tokens.
The Re‑learning Horror Story
Agents forget.
You pay.
They don't remember prior messages.
Each step starts cold.
Every message re‑sends history, code is re‑processed and re‑purchased. Even small codebases burn huge amounts of tokens.
The Re-learning Trap
Small tasks burn
10k–100k+ tokens
just to “learn” your code—before work starts
Then the burn continues, message after message.
At frontier prices, that's dollars per step.*
Rough estimate; your model and prompts will vary. The point: initial context dominates cost.
Stop the token burn
The re‑learning trap wastes money
and time you could spend building.
Open Source First:
We build on open source, so we give back. We'll make as much as possible open and free.
We'll also offer hosted and tailored cloud services (so we can eat).
Coming soon
CodeWeaver
A next‑gen model context protocol (MCP) server that composes and delivers just‑enough context.
- Language: Python
- License: MIT or Apache-2.0
- Goal: Give agents exactly what they need. Not more. Not less.
- Other cool stuff:
- A context delivery platform — tailor any context for any agent.
- Hybrid search, semantic ranking, 12+ embedding providers, auto‑indexing & codebase monitoring.
Coming soon
Thread
A codebase intelligence tool that keeps real‑time, semantically aware context aligned with your code as it changes.
- Language: Rust
- License: AGPL 3.0+
- Goal: Continuously analyze and understand your codebase to provide AI agents always accurate context.
- Other cool stuff:
- Built on Tree‑sitter for semantic parsing across 30+ languages.
- Graph‑based modeling to understand codebase‑wide relationships efficiently.