The open routing layer
Any model, any provider, any route. One familiar API for your next generation of agents. Know where every call goes and what it costs, down to the receipt.
The routes, the receipts, the balance, the agent rulebook, the host network and the privacy routes, with the limits of each.
Budget caps per request, hour, day and week.
Single-use ask-first approval window on /agents.
Deposit required to host. Hosts join by hardware attestation and probation.
Validity of a router-signed track-record certificate with a fresh pseudonym.
Routing, payments, proof and privacy in one layer. Keep your SDK; get a receipt, a budget and a choice of how to pay.
Any model, any provider, any route. One familiar API for your next generation of agents. Know where every call goes and what it costs, down to the receipt.
Keep the clients and request shapes your team already knows. Change the base URL and key, with OpenRouter-compatible routing preferences built in.
One balance for every model and provider. Fund it with USDC on Solana from your own wallet.
See the model, provider, tokens and cost behind each generation. A signed receipt travels with the response, and anyone can check it on /verify.
A request can begin without an account. Receive a 402 quote, pay, then retry with a payment proof.
Cap spending per request, hour, day and week. Choose models, lanes, tools and working hours. Ask first, or stop the agent before its next request. Manage your agents.
Opt in to public profiles with random slugs, an owner-chosen rulebook summary and the latest valid certificate. A2A-style cards and an MCP directory support discovery. Browse the directory or read profile limits.
Build and publish your agent sidecar image with the sealed recipe. The router verifies a registered agent’s TDX quote. No sealed agent is registered at trellis.example yet. Read the hosting limits.
Link Telegram from /agents with a one-time code. Approve or deny the same single-use requests and receive alerts through Trellis’s bot. Approval details pass through Telegram. Link your account.
End-to-end encrypted chat runs through the attested gateway. The router forwards ciphertext on this path. Other requests are read in memory to route them. Read how it works.
Run the approved build in a supported Intel TDX confidential VM. Fresh attestation and host policy govern admission. Join the network, read live network statistics or inspect public host records.
Filter by price, latency, quantization and data policy. Unhealthy providers leave the pool, with fallbacks built in.
Keep usage visible across your existing systems with OpenTelemetry export and usage records.
Per-agent ledgers export signed receipts as CSV or JSON. Router-signed certificates use a fresh pseudonym and last seven days. They are verifiable, not anonymous.
Roles, budgets and access for the people building with you: one routing layer for a shared workspace.
The same request shape reaches a standard model, an attested host or a Stock Token payment. Encrypted chat uses on-device encryption through the attested gateway; ordinary requests are read in memory by the router.
The same request shape reaches a standard model or an attested host. Requests are read in memory by the router to route them.
const res = await client.chat.completions.create({
model: "meta-llama/llama-3.3-70b-instruct",
messages,
provider: { sort: "price", allow_fallbacks: true },
});
res.provider; // who served it
res.usage.cost; // what it cost, in USDC
res.receipt.sig; // Ed25519 signaturePoint your client at the router, keep your code, and every call comes back with the route it took and the proof of what it cost.
Use the familiar API request shape. Set your model, provider preferences and budget.
model · provider{} · budgetThe router filters providers by health, price, data policy and your requirements, with fallbacks ready.
1/price² × uptime × qualityEvery generation returns itemized usage, cost and a signed receipt you can check on /verify.
ed25519 · sha-256 · /verifyThe Trellis Network is open for early hosts running the approved build. Admission checks the hardware and policy before a host serves requests.
Early hosts run the approved Intel TDX build in a supported confidential VM. Join with one command using join.mjs.
Fresh quote, signed host policy v1 in the key log, and sanctions screening of operator and payout addresses. New hosts start on probation with a public record.
Set boundaries on the key and attach an agent rulebook. Trellis’s router enforces it for requests through Trellis only.
Keep your SDK, choose providers and configure fallbacks. Cache and guardrails are opt-in.
Set request/hour/day/week caps, models, lanes, tools and working hours. Stop refuses the next request; the owner resumes.
Approve once on /agents within 15 minutes. Read signed receipts in the per-agent ledger and export CSV or JSON.
Follow the /agents alert feed, spend-alert webhook or Telegram via Trellis’s bot. Circuit breakers and progressive autonomy support spending caps up to 10x.
{ "name": "support-agent", "limit": 250, "limit_reset": "monthly", "rpm": 600, "tpm": 400000, "allowed_models": ["qwen/qwen3-32b", "deepseek/deepseek-r1"], "team": null, "pay_with_default": null, "expires_at": null, "guardrails": null, "routing": { "provider": { "sort": "latency", "allow_fallbacks": true } }}A private request goes only to a provider whose TEE evidence was verified minutes ago. If the evidence is missing or stale, the request fails closed.
The client encrypts on your device; the attested gateway handles inference while the router forwards ciphertext.
Trellis’s router reads request text in memory to route it. Attested serving alone does not hide prompts from the router.
Inspect signed receipts, the key transparency log and the data inventory. Receipts store hashes, token counts and cost, not prompt or answer text.
Check what the router has verified about any provider · Encrypted chat · What we keep
Encrypted chat, the agent rulebook and early-host admission are live at trellis.example. Sealed agent hosting is available, but no sealed agent is registered at trellis.example yet. Network host payouts and bond slashing are not switched on yet.
This router runs the core launch surface: chat, the model catalogue, API keys, prepaid credits, x402 per call and signed receipts. The areas below are built, and each card says whether this router has it enabled.
On-device encryption through the attested gateway, signed receipts, key transparency and per-host receipt anchoring.
Budgets, rules, approvals on /agents and Telegram, receipts, alerts, circuit breakers, progressive autonomy, certificates and opt-in profiles.
Open for early hosts on the approved Intel TDX build, automatic admission, probation, public host records and live network statistics.
Private RAG/files, PDF support, Chat, Tor onion access and blind tokens.
Agreements between agents are live: escrow and dispute contracts on Solana, with disputes ruled by a jury of models on attested hardware.
Trellis is an open inference router for Solana. A familiar API is the starting point. Clear payments, provider accountability and portable receipts are the reason to go further.
A route should tell you more than whether a request succeeded.
Which provider served it? What did it cost? What evidence supports its privacy claims? Who gets paid? Routing, USDC settlement, signed receipts and attestation checks run in the router, with contracts on Solana.
Keep your existing request shape. Point your client at your router’s /api/v1 and use a Trellis key. Streaming, tools, structured output and provider preferences work as you expect.
// Change two lines: the base URL and the key.
const client = new OpenAI({
baseURL: process.env.TRELLIS_BASE_URL, // "<your router>/api/v1"
apiKey: process.env.TRELLIS_API_KEY, // "sk-tr-v1-…"
});
const generation = await client.chat.completions.create({
model: "meta-llama/llama-3.3-70b-instruct",
messages: [{ role: "user", content: "Hello, Trellis." }],
});