Claude access, Ollama, more models & context7

1. Claude access
2. Model tiers per activity (subagent routing)
3. Ollama (local, no API key)
4. Adding more / cloud models
5. context7 — live library docs

---

Claude access

.aiflow/config.json → claude.auth (both supported; OAuth wins if both are set):

• apikey → ANTHROPIC_API_KEY (pay-per-use, console.anthropic.com).
• oauth → run claude setup-token → CLAUDE_CODE_OAUTH_TOKEN (uses your Claude plan).

Both live in .env (gitignored, never global).

Model tiers per activity (subagent routing)

Thinking-heavy work gets a strong model; mechanical work does not. .aiflow/config.json → modelRouting (default on) stamps the right model into each subagent’s frontmatter whenever aiflow apply runs — Claude Code only, no external router involved.

Tier	Default model	Subagents	Why
reasoning	opus (set fable if you prefer)	architect, planner, reviewer, security-advisor, requirements-check, modernization-advisor, orchestrator	Architecture, concepts, decomposition, review and security analysis are where a wrong call is expensive and long-lived.
implementation	sonnet	implementer, tester, quality-check, accessibility-checker	Bounded work with explicit acceptance criteria to check against.
mechanical	haiku	docs-sync, test-gap-advisor, dependency-auditor, performance-advisor, onboarder	Pattern-matching and enumeration — CI-grade checks, no design judgement.

Override the model per tier, or move a single agent to another tier:

{
  "modelRouting": {
    "enabled": true,
    "tiers":  { "reasoning": "fable", "implementation": "sonnet", "mechanical": "haiku" },
    "agents": { "implementer": "reasoning" }   // this one agent gets the reasoning model
  }
}

With "enabled": false every subagent runs on the session’s model and the model: lines are stripped again. Toggle with aiflow change-settings; tiers/agents are hand-edited and survive that (they are carried over, not rebuilt).

Escalate, never silently downgrade: if a “simple” task turns out to touch the architecture, it is architecture work — move it up a tier. Copilot and Codex have no subagents; pick the equivalent model manually and keep it stable within a thread.

Ollama (local, no API key)

Enable at aiflow init, or manage any time:

aiflow ollama add qwen3-coder     # add a model to config + pull it
aiflow ollama pull                # pull every model listed in config
aiflow ollama list                # what's installed

qwen3-coder (newest Qwen) is the recommended default. Selected models are written into .aiflow/router-config.json as a provider, so they’re actually used for easy/background work:

aiflow shell --router             # routes cheap/background steps to local models

Adding more / cloud models

For DeepSeek, OpenRouter, Gemini, and other providers:

1. Add the provider + key to ~/.claude-code-router/config.json (never committed).
2. Enable router in .aiflow/config.json.
3. Optional keys can also live in .env (DEEPSEEK_API_KEY, OPENROUTER_API_KEY, GEMINI_API_KEY).

Route trivial/background steps to cheap models; keep top Claude models for hard reasoning. Measure the effect with aiflow cost.

context7 — live library docs

context7 is an MCP server that fetches live, version-correct documentation for the libraries you use, so the agent codes against the real current API instead of a stale memory. Enabled by default (mcp.context7).

• In a session, just ask normally (“use the latest zod schema API”) — the agent calls context7 to pull current docs. You can also nudge it: “check context7 for the current Prisma migrate API”.
• Works keyless; a CONTEXT7_API_KEY in .env raises rate limits.
• Pair it with the code indexes: context7 = external library docs, graphify/cocoindex = your code. See Memory.