Use Roo Code with GammaInfra

Roo Code (the Cline fork with extended agent features) runs aggressive autonomous loops. Auto-approve mode makes those loops longer, which makes provider rate-limits hit harder. Routing through GammaInfra adds per-tool-call cost visibility and an automatic fallback chain so a 30-step task doesn't die mid-flight when one provider throttles.

The pain Roo Code users hit

Long auto-approve sessions burn tokens fast. Roo's auto-approve mode means the agent runs uninterrupted across 20–40 tool calls. Cost shape stays opaque until the provider invoice arrives.
Rate-limit halts kill auto-approve. When OpenAI or Anthropic throttles a key mid-session, Roo's auto-approve loop stops. The longer the session, the more annoying the restart.
Quality-tier upgrades require config edits. Switching from Sonnet to Opus for hard tasks means re-opening settings.

What changes with GammaInfra

Same smart-router drop-in as Cline (Roo Code is a Cline fork — the config flow is identical, on top of the OpenAI SDK shape). After the swap:

Every tool-call response carries X-GammaInfra-Cost-USD, summed in the dashboard per session.
Provider rate-limits cascade. The fallback chain serves the next provider transparently; Roo sees a 200 and keeps going.
Tool-call ID translation (toolu_* ↔ call_*) means the loop round-trips correctly when the cascade lands on Anthropic vs OpenAI.

Setup

1. Get a GammaInfra API key

Sign up at gammainfra.com and verify your email. $3 trial credit covers roughly one full Roo Code auto-approve task end-to-end.

2. Open Roo Code's settings

Open the Roo Code panel in VS Code (sidebar)
Click the settings gear icon in the top-right
Under API Provider, select OpenAI Compatible

3. Set base URL, key, and model

Base URL:    https://api.gammainfra.com/v1
API Key:     sk-gammainfra-...   (paste your GammaInfra key)
Model ID:    anthropic/claude-opus-4-7   (recommended for agent work)

4. Pick the right model

Roo Code benefits more from quality models than Cline because of the longer auto-approve sessions. Some patterns:

anthropic/claude-opus-4-7 — Anthropic flagship. Best for long auto-approve sessions where one bad decision early derails the whole run.
anthropic/claude-sonnet-4-6 — Sonnet. Solid quality/cost trade for routine tasks.
openai/gpt-5 — OpenAI flagship with strong tool-use support.
deepseek/deepseek-v4-pro — cheaper reasoning option with thinking mode enabled by default.
gammainfra/auto — task-aware routing. Good when you're not sure which tier fits.

Auto-approve safety: GammaInfra's fallback chain returns a 200 to Roo Code even when the cascade went through 2–3 providers. Roo's auto-approve mode won't notice or pause. If you want auto-approve to stop on degraded routes, watch the X-GammaInfra-Attempted-Count response header in Roo's logs and add manual checkpoints when it climbs.

Cost-conscious patterns

Max-latency budget header. X-GammaInfra-Max-Latency-Ms: 30000 as a custom header — instead of an auto-approve session quietly stalling on a slow provider response.
Cost-quality dial. X-GammaInfra-Cost-Quality: 0.5 as a custom header to balance cost and quality globally. Drop to 0.2 for sessions where quality matters more; push to 0.8 for routine cleanup runs.
BYOK for power users. Add your own Anthropic/OpenAI provider keys via the dashboard's Provider Keys tab. Roo Code traffic routes via your keys (BYOK mode) at 1–2% per request.

Verify it's working

Start a Roo Code auto-approve task. Then in the dashboard:

Recent requests appear with timestamps, resolved provider/model, latency, and cost.
Daily roll-ups show total spend by model. Drill into a heavy auto-approve day to see the per-tool-call shape.
The request-detail view shows the fallback chain if GammaInfra cascaded providers.

Cline / Roo Code differences

For most setup decisions, the answer is "same as Cline." The only differences worth noting for GammaInfra users:

Roo's auto-approve mode means longer sessions ⟹ more fallback opportunities ⟹ more value from GammaInfra's chain.
Roo's extended agent features (additional tools, longer context) lean harder on quality tiers — pin Opus/Sonnet rather than relying solely on gammainfra/auto.

The full setup walkthrough is also applicable from the Cline integration page.

Trade-offs

Latency. ~10–50 ms overhead per request. Negligible during agent thinking gaps.
Cost. 3% top-up fee (launch window) / 5% standard on managed credits. Pass-through provider rates on tokens. BYOK 1–2% per request alternative.
Privacy. Prompts and responses aren't logged by default. Privacy policy.

Ready to try it?

Get a GammaInfra API key →

$3 free trial credit on signup, $10 minimum top-up. Pass-through provider token rates plus 3% top-up fee during the launch window.

Frequently asked questions

How do I configure Roo Code to use GammaInfra?

Same as Cline (Roo Code is a Cline fork). Open Roo Code Settings, choose OpenAI Compatible as the provider, set the Base URL to https://api.gammainfra.com/v1, paste your sk-gammainfra-... key, and pick a model like gammainfra/auto. Roo's agent loops and tool calling work end-to-end.

Does Roo Code's agent mode work with smart routing?

Yes. Roo's agent loop fires multiple steps per task (read files, plan, edit, test). The router classifies each step independently — file-reads typically route to extraction-tier models, code-generation steps to code-tier models, planning steps to reasoning-tier. This typically reduces total run cost vs running every step through one flagship.

Can I see per-step cost in Roo's task log?

Roo Code's task log doesn't surface response headers directly. Use the GammaInfra dashboard to see per-step costs. Each agent step is a separate request_logs entry — filter by API key and time window to isolate one Roo task.

What if Roo's tool call format differs between providers?

GammaInfra translates tool_call.id between Anthropic's toolu_* format and OpenAI's call_* format at the wire boundary. As long as Roo's session stays on one provider, IDs round-trip cleanly. Switching providers mid-conversation (e.g. fallback fired) can break tool-call continuity because each provider validates IDs it issued — design Roo workflows to commit to one provider per task and switch between tasks.

Does Roo Code benefit from hedged requests?

Use gammainfra/fast as Roo's model and (if your operator has enabled hedging) the router fires two providers in parallel and takes the first success. For Roo's iterative agent loops, this is most useful on the planning / reasoning steps where latency variance hurts. Skip it for high-volume tool-execution steps where the 2× cost outweighs the latency win.