Tool landscape
What was looked at around the project, and where each piece ended up. The planned comparison matrix was not filled: the choices were made on benchmark runs instead.
Gateways
| Tool | What it is | Status in llm_brain |
|---|---|---|
| OpenRouter | hosted aggregator, both dialects, per-key limits | the upstream of every tier |
| LiteLLM | general-purpose proxy/gateway | not used (why) |
| claude-code-router | proxy from Claude Code to other models | not used; reference if a translator is ever needed |
Why not LiteLLM
The first plan was LiteLLM for a week in Phase 0, then a native proxy. It
was dropped once OpenRouter turned out to speak both dialects and to offer
per-key daily limits without a database: Phase 0 ran on OpenRouter
directly (Stack). As a gateway LiteLLM enforces budgets only
with Postgres, is one more service to run, routes by config only, and
doesn't reuse llm_brain's usage DB, board or brain/auto. If a format
translator is ever needed, its Anthropic ↔ OpenAI mapping is a reference
to port, alongside claude-code-router.
CLIs / IDEs
| Tool | Status |
|---|---|
| Claude Code | daily agent, through the proxy (Which CLI) |
| aider | cheap editor (architect/editor) |
| OpenCode | benchmarked (brain bench run --tool opencode) |
| Cline / Roo Code, Continue | not evaluated; any of them works as a client of the OpenAI-compatible endpoint |
Related repos
- claude-kit — portable hooks/skills, a submodule of this repo
- agent-orchestrator — a client (
agoprofile) - AgentsBoard — macOS mission control for Claude Code/Codex: a potential client of the layer