Add Provider : agentrouter.org #2784

Activity

1. github-actions commented on Sep 25, 2025

   

   github-actionsBot

   on Sep 25, 2025 – with GitHub Actions

   Contributor

   More actions

   This issue might be a duplicate of existing issues. Please check:

   • In latest version v0.11.3 How to config and use GLM-4.5 model if there isn't in the list? #2760: Similar question about configuring GLM-4.5 model when it's not in the available models list

   Feel free to ignore if none of these address your specific case.
2. rekram1-node

   assigned

   fwang

   on Sep 25, 2025
3. rekram1-node commented on Sep 25, 2025

   

   rekram1-node

   on Sep 25, 2025

   Last edited by rekram1-node

   Contributor

   More actions

   We will need to add to models.dev, you should still be able to use it if you define custom provider tho: https://opencode.ai/docs/providers/#custom-provider
4. santoshd commented on Oct 15, 2025

   

   santoshd

   on Oct 15, 2025

   More actions

   I tried working with Aiden on discord to add the service through a custom provider. I suspect there's some filtering on the part of Agentrouter that prevents non-approved tool use. They have listed codex, claude as supported. I have attempted another CLI tool as well which also fails. I've written to the Agentrouter team and they have not responded.
5. alpgul commented on Nov 28, 2025

   

   alpgul

   on Nov 28, 2025

   More actions

   Temporary solution:

   "agentrouter": {
         "npm": "@ai-sdk/openai-compatible",
         "name": "agentrouter",
         "options": {
           "baseURL": "https://agentrouter.org/v1",
           "headers": {
             "Authorization": "Bearer apiKey",
             "User-Agent": "RooCode/3.34.8",
             "X-Title": "Roo Code",
             "HTTP-Referer": "https://github.com/RooVetGit/Roo-Cline",
             "X-Stainless-Runtime-Version": "v22.20.0",
             "X-Stainless-Runtime": "node",
             "X-Stainless-Arch": "x64",
             "X-Stainless-OS": "Linux",
             "X-Stainless-Lang": "js"
           }
         },
         "models": {
           "glm-4.6": {
             "name": "glm-4.6",
             "limit": {
               "context": 200000,
               "output": 65536
             }
           }
         }
       }

   👍React with 👍5Reacted by Dima, TudeStillLearning!, M. AZZUHRY, Ade Peruzi Septiawan and Anatoliy D. Uvarov
6. thdxr

   closed this as completedon Mar 25, 2026
7. gethoopp commented on Apr 23, 2026

   

   gethoopp

   on Apr 23, 2026

   More actions

   unauthorized client detected, contact support for assistance at https://discord.com/invite/V6kaP6Rg44

   got this when use api key from agent router
8. Lypt1x commented on Apr 23, 2026

   

   Lypt1x

   on Apr 23, 2026

   Last edited by Lypt1x

   More actions

   @gethoopp

   unauthorized client detected, contact support for assistance at https://discord.com/invite/V6kaP6Rg44

   got this when use api key from agent router

   Got this working on opencode 1.14.22. Three separate problems stacking on top of each other:

   1. AgentRouter fingerprints the HTTP client. Their backend is OneAPI-based (you can see X-Oneapi-Request-Id in response headers), and unless the request looks like Claude Code they reject it with unauthorized client detected. A spoofed User-Agent alone isn't enough, you also need the x-stainless-* set and anthropic-* headers that the real Claude Code CLI sends.

   2. opencode bug #22608: custom headers in opencode.json (especially User-Agent) are silently dropped before fetch. So any config-based approach (like the RooCode one posted above) can't work regardless of what you put in there.

   3. AgentRouter emits a malformed data: null SSE chunk inside streams, which makes the AI SDK's Zod validator throw Type validation failed: Value: null has to be filtered.

   Workaround: a small plugin that patches globalThis.fetch for agentrouter.org only while other providers are untouched.

   ~/.config/opencode/opencode.json

   {
     "$schema": "https://opencode.ai/config.json",
     "provider": {
       "agentrouter": {
         "npm": "@ai-sdk/openai-compatible",
         "name": "agentrouter",
         "options": {
           "baseURL": "https://agentrouter.org/v1",
           "apiKey": "sk-YOUR_KEY"
         },
         "models": {
           "glm-4.6":                   { "name": "glm-4.6",                   "limit": { "context": 200000, "output": 65536 } },
           "claude-opus-4-6":           { "name": "claude-opus-4-6",           "limit": { "context": 200000, "output": 65536 } },
           "claude-haiku-4-5-20251001": { "name": "claude-haiku-4-5-20251001", "limit": { "context": 200000, "output": 65536 } },
           "deepseek-v3.1":             { "name": "deepseek-v3.1",             "limit": { "context": 200000, "output": 65536 } }
         }
       }
     }
   }

   No headers block, the plugin does that. apiKey handles Authorization: Bearer automatically via the SDK.

   Notice: I may be wrong with the upstream limits. Needs to be checked.

   ~/.config/opencode/plugins/agentrouter-headers.js

   // Scoped strictly to agentrouter.org. Other providers unaffected.
   const AGENTROUTER_HOST = "agentrouter.org";
   
   const INJECTED_HEADERS = {
     "User-Agent": "claude-cli/1.0.108 (external, cli)",
     "anthropic-version": "2023-06-01",
     "anthropic-beta": "claude-code-20250219,oauth-2025-04-20",
     "anthropic-dangerous-direct-browser-access": "true",
     "x-app": "cli",
     "x-stainless-lang": "js",
     "x-stainless-package-version": "0.55.1",
     "x-stainless-os": "Windows",
     "x-stainless-arch": "x64",
     "x-stainless-runtime": "node",
     "x-stainless-runtime-version": "v22.0.0",
   };
   
   function isAgentRouterUrl(url) {
     try {
       const p = typeof url === "string" ? new URL(url) : new URL(url.url ?? url.href ?? String(url));
       return p.hostname.endsWith(AGENTROUTER_HOST);
     } catch { return false; }
   }
   
   // Strip malformed `data: null` SSE chunks that crash the AI SDK Zod validator.
   function sanitizeSseBody(body) {
     if (!body) return body;
     const dec = new TextDecoder();
     const enc = new TextEncoder();
     let carry = "";
     const t = new TransformStream({
       transform(chunk, controller) {
         const text = carry + dec.decode(chunk, { stream: true });
         const nl = text.lastIndexOf("\n");
         if (nl === -1) { carry = text; return; }
         const complete = text.slice(0, nl + 1);
         carry = text.slice(nl + 1);
         const filtered = complete.split("\n")
           .filter(l => l.trim() !== "data: null" && l.trim() !== "data:null")
           .join("\n");
         if (filtered.length) controller.enqueue(enc.encode(filtered));
       },
       flush(controller) {
         if (carry.trim() !== "data: null" && carry.trim() !== "data:null" && carry.length) {
           controller.enqueue(enc.encode(carry));
         }
       },
     });
     return body.pipeThrough(t);
   }
   
   function patchFetchOnce() {
     if (globalThis.__agentrouter_fetch_patched__) return;
     globalThis.__agentrouter_fetch_patched__ = true;
     const orig = globalThis.fetch.bind(globalThis);
   
     globalThis.fetch = async function (input, init) {
       try {
         const url = typeof input === "string" || input instanceof URL ? input : input?.url;
         if (url && isAgentRouterUrl(url)) {
           const headers = new Headers(init?.headers ?? (input && input.headers) ?? undefined);
           for (const [k, v] of Object.entries(INJECTED_HEADERS)) headers.set(k, v);
           const newInit = { ...(init ?? {}), headers };
           const res = input instanceof Request
             ? await orig(new Request(input, newInit))
             : await orig(input, newInit);
           const ct = res.headers.get("content-type") ?? "";
           if (res.ok && ct.includes("text/event-stream") && res.body) {
             return new Response(sanitizeSseBody(res.body), {
               status: res.status, statusText: res.statusText, headers: res.headers,
             });
           }
           return res;
         }
       } catch {}
       return orig(input, init);
     };
   }
   
   export const AgentRouterHeadersPlugin = async ({ client }) => {
     patchFetchOnce();
     try {
       await client.app.log({
         body: { service: "agentrouter-headers", level: "info", message: "fetch patched for agentrouter.org" },
       });
     } catch {}
     return {};
   };

   Fully quit and restart opencode after adding this (config + plugins load at startup).

   Heads up on title generation

   Opencode uses small_model for background title generation. When that goes through AgentRouter you'll occasionally get sensitive_words_detected from their Chineeese content filter on otherwise normal prompts, the actual chat works fine, only the auto-title fails.

   Workarounds:

   Route titles elsewhere:

   { "model": "agentrouter/claude-opus-4-6", "small_model": "anthropic/claude-haiku-4-5" }

   Or force titles onto the main model:

   { "model": "agentrouter/claude-opus-4-6", "small_model": "agentrouter/claude-opus-4-6" }

   Would obviously be nicer if opencode just fixed #22608 so you could configure the headers directly in opencode.json, but for now this does the job.

   Hope it helps, happy coding!

   👍React with 👍8Reacted by dabbrahost, Adriana Eka Prayudha, Cycle, Arnab Mondal, george3232-foo, sbeygi, a1c0d3r and Haliim Pamungkas

   👎React with 👎1Reacted by Void

   😄React with 😄3Reacted by Nix Seymour, Ujang Karnadi and Arda Kılıçdağı
9. sbeygi commented on Jun 26, 2026

   

   sbeygi

   on Jun 26, 2026

   Last edited by sbeygi

   More actions

   "agentrouter": {
     "npm": "@ai-sdk/openai-compatible",
     "name": "agentrouter",
     "options": {
       "baseURL": "https://agentrouter.org/v1",
       "apiKey": "YOUR-API-KEY"
     },
     "models": {
       "glm-5.2": {
         "name": "glm-5.2",
         "limit": { "context": 1000000, "output": 1000000 }
       },
       "claude-opus-4-6": {
         "name": "claude-opus-4-6",
         "limit": { "context": 1000000, "output": 1000000 }
       },
       "claude-opus-4-7": {
         "name": "claude-opus-4-7",
         "limit": { "context": 1000000, "output": 1000000 }
       },
       "claude-opus-4-8": {
         "name": "claude-opus-4-8",
         "limit": { "context": 1000000, "output": 1000000 }
       }
     }
   }
   

   The solution offered by Lypt1x works. This is just an updated JSON for opencode.

   👍React with 👍1Reacted by a1c0d3r