Trilok Labs · est. 2024

trilok

A harness for evolving agents.

Seven open-source libraries that make the loop tight enough that agents can iterate themselves. Tools become the environment. Traces become genomes.

RustZigPythonWebAssembly

What we believe

The model is the seed.
The harness is the soil,
the sunlight, and the pressure
that decides whether anything grows.

Trilok is the substrate underneath. Tight loops, schema-first tool calls, trajectory recording that lets the next iteration learn from the last.

Seven organs · one harness

The libraries.

Each one solves a real bottleneck in the loop. Each one ships separately. Together they’re the soil under everything we build.

File Reader
Muonry
Drop-in MCP upgrade for Claude Code, Codex, Cursor, Windsurf. Stops agents burning weekly caps on full-file reads.
40× Leaner0.7ms / op344M tokens saved
Read source
Code Search
CodeDB
Code intelligence for agents. Trigram v2 with an O(1) inverted word index. Search is the agent’s first sense organ.
538× vs ripgrep16 MCP toolsZero deps
Read source
Package Manager
Nanobrew
Drop-in Homebrew + apt-get replacement, in Zig. The slow part inside the sandbox wasn’t compute — it was apt-get.
13× Warm Install3ms InstallsWritten in Zig
Read source
Browser
Kuri
The agent’s web arm. Single Zig binary speaking Chrome DevTools Protocol. No Node, no npm.
7.3× LeanerZig + CDPNo Node
Read source
Orchestrator
DevSwarm
MCP server in Zig. One agent spawns a team — orchestrator splits the task, workers run in parallel, synthesiser stitches.
Parallel AgentsZig MCPTelemetry → Graff
Read source
Web Framework
merjs
Next.js-style web framework in Zig. Zero Node.js. File-based routing, SSR, WASM client logic, Cloudflare Workers.
<5ms Cold StartZig + WASMNo node_modules
Read source
Python Framework
TurboAPI
FastAPI-compatible Python with a Zig HTTP core. Drop-in syntax, dramatically faster wire path.
21× vs FastAPIPython + ZigDrop-in
Read source

Receipts

Numbers, not vibes.

538×
CodeDB
vs ripgrep · pre-indexed queries
40×
Muonry
outline vs full file read · per op
21×
TurboAPI
vs FastAPI · ASGI throughput
3ms
Nanobrew
warm package install on macOS

Each number reproducible · benchmarks on GitHub

Seven libraries · one harness

Same algorithm as biology.
Different substrate.

2026 is another year that everything handcrafted gets replaced by everything learnt. We’d rather build the soil than another seed.

Browse on GitHubRead the notes