Skip to content

ROADMAP.md

Latest commit

 

History

History
220 lines (163 loc) · 6.97 KB

ROADMAP.md

File metadata and controls

220 lines (163 loc) · 6.97 KB

brcc Roadmap

Vision

brcc = Claude Code + any model + no rate limits + pay-per-use USDC

The goal is to let any developer run Claude Code with any LLM model, selecting and switching models freely, paying only for what they use.

Current State (v0.5.0)

User runs: brcc start --model openai/gpt-5.4

  brcc proxy (localhost:8402)
    ├── Receives Anthropic-format requests from Claude Code
    ├── Overrides model name to user's choice
    ├── Signs x402 payment with user's wallet
    └── Forwards to blockrun.ai/api/v1/messages

Working:

  • brcc setup base|solana — create wallet
  • brcc start --model <model> — start proxy + launch Claude Code
  • brcc models — list 50+ models with pricing
  • brcc balance — check USDC balance
  • Dual chain support (Base + Solana)
  • Free model (nvidia/nemotron-ultra-253b) tested end-to-end

Limitations:

  • Cannot switch models inside Claude Code — must restart with different --model
  • Claude Code's /model only shows Anthropic models
  • Auth conflict warning when user has existing claude.ai login

Phase 1: Dynamic Model Selection (next)

Goal

Let users switch between any BlockRun model inside Claude Code without restarting.

Approach: Combine tweakcc + claude-code-router patterns

From tweakcc:

  • allowCustomAgentModels patch — removes Claude Code's model name validation
  • Enables arbitrary model names in Claude Code's /model picker
  • Patches Claude Code's minified JS directly

From claude-code-router:

  • Dynamic /model provider,model_name syntax inside Claude Code
  • Smart routing based on task type (simple/complex/reasoning/long-context)
  • Transformer chain for format conversion between providers
  • Per-subagent model override via tags

Implementation Plan

1. Add tweakcc patching to brcc

brcc patch          # Patch Claude Code to allow custom model names
brcc patch --undo   # Restore original Claude Code

Key patches to apply:

  • allowCustomAgentModels — unlock arbitrary model names
  • contextLimit override — support models with different context windows
  • subagentModels — set different models for plan/explore/general-purpose agents

Reference: /Users/vickyfu/tmp/tweakcc/src/patches/

2. Smart model routing in proxy

Borrow from ClawRouter's 14-dimension classifier:

Claude Code request → brcc proxy analyzes request →
  Simple task (definitions, math) → cheapest model (DeepSeek, NVIDIA free)
  Code task (editing, debugging) → code-optimized model (Claude Sonnet, GPT-5)
  Reasoning task (proofs, planning) → reasoning model (o3, Grok reasoning)
  Long context (large files) → large context model (GPT-5.4 1M, Gemini)

User can override with explicit /model selection.

Reference: /Users/vickyfu/Documents/blockrun-web/ClawRouter/

3. Env var model mapping

brcc start sets these env vars so Claude Code's built-in /model picker maps to BlockRun models:

ANTHROPIC_DEFAULT_SONNET_MODEL=anthropic/claude-sonnet-4.6
ANTHROPIC_DEFAULT_OPUS_MODEL=anthropic/claude-opus-4.6
ANTHROPIC_DEFAULT_HAIKU_MODEL=deepseek/deepseek-chat  # cheap alternative
CLAUDE_CODE_SUBAGENT_MODEL=anthropic/claude-haiku-4.5

User can customize in ~/.blockrun/brcc-config.json.

4. Transformer chain

For non-Anthropic models, the proxy needs format conversion:

Claude Code (Anthropic format)
  → brcc proxy
    → If Anthropic model: pass through to /v1/messages
    → If OpenAI model: convert Anthropic→OpenAI format, call /v1/chat/completions
    → Convert response back to Anthropic format
  → Claude Code

BlockRun's /v1/messages endpoint already handles this server-side, so the proxy just forwards. But for edge cases (streaming, tool calling), client-side transformation may be needed.

Phase 2: Installation UX

Goal

One-line install that sets up everything.

Implementation

# Install brcc + Claude Code + patch in one command
curl -fsSL https://brcc.blockrun.ai/install.sh | bash

The install script:

  1. Install Node.js if missing
  2. Install Claude Code: curl -fsSL https://claude.ai/install.sh | bash
  3. Install brcc: npm install -g @blockrun/cc
  4. Create wallet: brcc setup base
  5. Patch Claude Code: brcc patch
  6. Show wallet address + funding instructions

Uninstall

brcc patch --undo    # Restore Claude Code
npm uninstall -g @blockrun/cc

Phase 3: Smart Defaults

Goal

Zero-config experience — brcc auto-picks the best model for each task.

Implementation

brcc start --smart   # Auto-route every request to optimal model

Uses ClawRouter's classifier to analyze each request:

  • Token count, code presence, reasoning markers, creative markers
  • Routes to cheapest capable model
  • User sets budget: brcc start --smart --budget 0.01 (max $0.01 per request)

Phase 4: Team Features

Goal

Teams share a wallet and track per-developer usage.

Implementation

brcc team create "My Team"
brcc team add dev@example.com
brcc team budget 100          # $100 monthly budget
brcc team usage                # Per-developer breakdown

Technical Reference

Key repos

Repo What to borrow
tweakcc JS patching engine, allowCustomAgentModels, subagentModels, prompt customization
claude-code-router Dynamic /model switching, transformer chain, smart routing by task type
ClawRouter 14-dimension request classifier, cost-optimized model selection
OpenRouter ANTHROPIC_DEFAULT_*_MODEL env vars, ANTHROPIC_AUTH_TOKEN pattern

Key Claude Code env vars

ANTHROPIC_BASE_URL              # API endpoint (brcc sets to localhost proxy)
ANTHROPIC_API_KEY               # API key (brcc sets dummy key)
ANTHROPIC_MODEL                 # Default model
ANTHROPIC_DEFAULT_OPUS_MODEL    # What /model opus resolves to
ANTHROPIC_DEFAULT_SONNET_MODEL  # What /model sonnet resolves to
ANTHROPIC_DEFAULT_HAIKU_MODEL   # What /model haiku resolves to
CLAUDE_CODE_SUBAGENT_MODEL      # Model for subagents
ANTHROPIC_CUSTOM_MODEL_OPTION   # Custom model in /model picker (no validation)

Architecture target

brcc start
  │
  ├── Patch Claude Code (tweakcc engine)
  │     └── Unlock custom model names
  │
  ├── Start proxy (localhost:8402)
  │     ├── Request classifier (ClawRouter)
  │     ├── Model router (CCR pattern)
  │     ├── Transformer chain (Anthropic↔OpenAI)
  │     ├── x402 payment signing
  │     └── Response streaming
  │
  └── Launch Claude Code
        ├── ANTHROPIC_BASE_URL → proxy
        ├── Model env vars → BlockRun models
        └── /model picker → all 50+ models