The whole
picture.
Connect your accounts. Order your routes. See which provider answered, what it used, and what happened along the way.
A limit is a turn.
When a provider hits a limit before output starts, Conflux can try the next eligible route. Here’s a request that did exactly that.
An answer, with a trail.
OpenRouter answered after Groq rate-limited. Both hops stayed in the same inspectable receipt.
Groq
openai/gpt-oss-120bOpenRouter
cohere/north-mini-code:freeA saved pre-output failover. This guided replay sends no API request.
The agent is
Conflux, too.
Start a project, describe the work, and build with Conflux’s own Rust harness. One coding loop, in your terminal or your console.
Meet the Conflux harness- Plan, build, verify.
Read files, make precise edits and run commands you approve.
- Set the effort. Keep control.
Choose a model, reasoning preference and request limits. Model support still matters.
- Pick up where you left off.
Opt into encrypted task history, then resume in the same project.
Linux preview · shell off by default · no separate coding CLI required
cargo build --release -p conflux-harness -j 2
target/release/conflux init ./my-project
target/release/conflux chat \
--project ./my-project \
--model conflux/code \
--effort high \
--saveSet your gateway key privately in CONFLUX_API_KEY first. Add --allow-shell only when you want to review host commands.
The gateway finds an eligible provider. The harness owns the project, tools and approvals.
Read the preview limitsKnow your
free allowance.
Choose your accounts and a typical request size. See a daily workload estimate built from published limits.
Build your daily estimate
Assumes unused allowances on your own eligible accounts. Each provider pool counts once.
Include system prompts, conversation history, tool results and the answer. A coding turn can make several requests.
How this estimate worksAcross 5 selected pools at 2,000 tokens per request.
Request caps can reduce usable capacity. OpenRouter contributes estimated request capacity only.
A planning scenario, not guaranteed tokens or a live balance. Model access, context limits, per-minute caps, availability and provider terms still apply.
Sources, assumptions and the other free providers
Official sources checked 2026-10-24. Model caps are not multiplied by keys, aliases or models. The selected pools total 1,000,000 documented token-cap units per day; the default scenario is about 940,000 workload tokens after request caps.
- LLM7 — Input + output per 24 hours. Shared across free models. Quota may be reduced without notice.
- Groq — One model, one organization. 30 RPM / 8K TPM also apply; your console can report different limits.
- SambaNova — Free account without a payment method. Use the lower of the token and request caps.
- xKiro API — No-card baseline. The 1M Telegram-verified tier and signup rewards are excluded.
- OpenRouter — Zero-purchase baseline: 50 requests/day, 20 RPM. This is a request cap, not a token grant.
Cloudflare includes 10,000 Neurons/day; Neurons are compute units. Mistral, NVIDIA, Gemini, Z.ai, ZenMux, Kilo and SiliconFlow have no verified universal daily token grant here. They are excluded from the numeric total, so unknown does not mean zero.
Signup credits, paid-plan allowances, native-only services and unverified offers are excluded. OpenRouter aliases and models such as Arcee share OpenRouter’s pool. A provider’s published offer can change.
Your gateway.
Your boundaries.
Decide how requests recover, give tools their own budgets, and see the limits behind your pool.
A lane for each tool.
Give Codex, Claude Code or OpenCode a dedicated key, model and token budget. Keep each tool’s usage visible.
See it
in action.
Installed coding tools. Actual provider responses. Watch the recorded task, then inspect its matching evidence.
The provider changed. The task continued.
Recorded evidence · demo credentials only
Edited live Claude Code recording with its matching gateway receipt. Groq returned HTTP 429; OpenRouter completed the request with HTTP 200. This proves rate-limit failover before output, not total account-credit exhaustion.
Keep your
coding setup.
A gateway URL, a key and a model. Use a Harness Profile to keep each tool’s configuration and budget together.
curl http://localhost:3000/v1/chat/completions \
-H "Authorization: Bearer $CONFLUX_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "conflux/chat",
"messages": [{
"role": "user",
"content": "Hello, Conflux."
}],
"max_tokens": 32
}'Chat Completions, plus practical Responses and Messages dialects. Check compatibility.
Before you
connect.
Three ways in. Start with the one that fits your work—and know what you’re giving access to.
A project and a provider pool.
Use the native Rust harness when you want Conflux to handle the coding task as well as the model connection.
- A Linux workspace for the harness
- A tool-capable route and your own gateway key
- A choice: review every edit, or allow project edits
Is this a free AI subscription?
Conflux is MIT software you host yourself. You bring official provider accounts; their allowances, model access and terms still apply. Paid connections are optional and require explicit consent.
Can it continue after a provider stops?
Before substantive output begins, the gateway can try the next eligible route. A committed stream is never silently replayed. Natural exhaustion across three accounts remains unverified.
Where do my code and keys go?
Files stay on the harness host. Relevant instructions, file contents and tool results are sent to your configured model provider. Provider keys are encrypted at rest; gateway keys are hashed. Task history is saved only when you opt in.
What does “releasing soon” mean?
The source preview is being prepared for public release. The native workflow has passed isolated tool and gateway tests; long live coding and universal provider/client acceptance remain open. No public package or hosted service is promised today.
Conflux is
releasing soon.
Explore the source preview, choose your setup, and see what is ready before you connect your accounts.
Releasing soon. The gateway and native Rust harness are in review. Explore the guides; broad live acceptance is still in progress.