API reference

Everything the gateway exposes, grouped by auth model. Inference endpoints take an API key (Authorization: Bearer br-…); dashboard endpoints use the browser session cookie set at sign-in and are listed here for completeness — most people use them through the dashboard UI. A machine-readable schema lives at openapi.json.

Inference (API key)

POST /v1/chat/completions

OpenAI-compatible chat completions, streaming and non-streaming. Accepts theBharatRouter extensions (optimize,optimize_weights, provider, data_policy,upstream_key, fallbacks, …) — they are consumed by the router and stripped before the request leaves the gateway. The provider that actually served the request is echoed in the x-br-provider response header.

curl https://api.bharatrouter.com/v1/chat/completions \
  -H "Authorization: Bearer br-..." -H "Content-Type: application/json" \
  -d '{
    "model": "qwen2.5-7b-instruct",
    "messages": [
      {"role": "system", "content": "You answer in Hindi."},
      {"role": "user", "content": "What is the capital of Maharashtra?"}
    ],
    "max_tokens": 200,
    "optimize": "uptime"
  }'

Response (non-streaming) is a standard chat-completion object with token usage:

{
  "id": "chatcmpl-...",
  "object": "chat.completion",
  "model": "qwen2.5-7b-instruct",
  "choices": [{ "index": 0, "message": { "role": "assistant", "content": "मुंबई ..." }, "finish_reason": "stop" }],
  "usage": { "prompt_tokens": 28, "completion_tokens": 41, "total_tokens": 69 }
}

With "stream": true the response is text/event-stream. The gateway injects stream_options.include_usage on providers that support it, so the final chunk before [DONE] carries the usage block.

POST /v1/messages

Anthropic-native Messages API surface — same wire format Claude Code and the Anthropic SDKs already speak, so you can repoint an existing Anthropic client at BharatRouter without rewriting to the OpenAI shape. Serves the Claude models (claude-sonnet-5, claude-opus-4.8, claude-haiku-4.5) with the same routing, residency and INR metering as the chat surface.

Three ways to present your br- gateway key, in precedence order — pick whichever your client makes easy:

# Anthropic SDK — one line changes (base_url), your br- key goes in api_key
from anthropic import Anthropic
client = Anthropic(base_url="https://api.bharatrouter.com", api_key="br-...")
client.messages.create(model="claude-sonnet-5", max_tokens=256,
  messages=[{"role": "user", "content": "Namaste!"}])

# curl, native format
curl https://api.bharatrouter.com/v1/messages \
  -H "x-api-key: br-..." -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{ "model": "claude-sonnet-5", "max_tokens": 256,
    "messages": [{"role": "user", "content": "Namaste!"}] }'

POST /v1/embeddings

OpenAI-compatible embeddings with the same extensions, routing, and failover as chat. India-resident models served on the platform key: bge-m3 (1024-dim).

curl https://api.bharatrouter.com/v1/embeddings \
  -H "Authorization: Bearer br-..." -H "Content-Type: application/json" \
  -d '{ "model": "bge-m3", "input": ["namaste", "vanakkam"], "data_policy": "india_only" }'

Catalog & health (public, no auth)

curl -s https://api.bharatrouter.com/v1/models | jq '.data[] | select(.id=="qwen2.5-7b-instruct")'

Discovery (public)

Dashboard API (session cookie)

Set by GitHub/Google OAuth at sign-in(GET /auth/:provider → callback → br_session cookie;POST /auth/logout clears it). Org owners can do everything; some endpoints are owner-only as marked.

Account

API keys

Full semantics on API keys & limits.

BYOK keys

Full guide on the BYOK page.

Billing

Full semantics on Credits & billing.

Routing chains

Saved org-wide fallback chains — full guide on Routing.

Collections

Full guide on Collections.

Custom endpoints (BYOE)

Full guide on Bring your own endpoint.

Reliability monitoring

Full guide on Reliability monitoring.

Teams & workspaces

Full guide on Teams & workspaces.

Usage & activity

Full guide on Usage, activity & rankings.

Coding agent (brcode CLI)

BharatRouter Code (brcode) runs a coding agent on GLM through the gateway. One org profile — default model, MCP connectors, and optional rules — lets a teammate inherit your whole setup with brcode login && brcode init.

Agent identity (Ekam)

Every machine running a coding agent gets an invisible per-machine agent principal — an Ekam-governed identity created once and reused thereafter, so spend, memory writes and audit trails attribute to the agent rather than the human's key. The claude-native installer wires this automatically via Claude Code's apiKeyHelper; toggle withbr-mode identity on|off. Admins see and revoke agents in the console.

Requests carrying an agent identity are attributed in GET /v1/usage'sby_agent split and stamped onto shared-memory writes viax-br-agent-id.

Shared memory (agents)

An org-level object store the gateway fronts for agent memory: journals, records and private notes live under your own S3/KOS bucket, scope-checked per key. Default-on for users — it activates the moment an admin configures the store; until then memory endpoints return404 code memory_not_configured. All keys are validated (no ..or absolute paths) and joined under the org's base_prefix server-side.

Store setup (admin)

Owner/admin only — via the console's Memory storage card or the API. Credentials are encrypted at rest (same envelope as BYOK keys) and never returned.

Scope model

PII posture: the personal users/<uid>/ scope is never readable via team paths; the pattern is extraction, not transcription — agents write distilled facts, not raw conversation; every object carries provenance metadata so erasure requests can be honoured.

Peer shares

Peer shares add an explicit, tenant-local collaboration scope without widening personal, workspace or org access. A safe space slug resolves toshared/<space>/. Owners always have access; each granted member gets read/write access. Org owners/admins may administer a share but do not gain data access unless they are its owner or a granted member. ACL changes require a full-scope, user-backed key.

Objects & journal

In Claude Code the memory plane surfaces as automatic SessionStart recall plus/br-remember (team journal), /br-remember-me (private note) and/br-memory status. Journal entries expire after journal_ttl_days; anything worth keeping is promoted — by a human — into records.

Webhooks

Response headers

Platform limits

Per-key limits and monthly ₹ budgets are configurable — seeAPI keys & limits. Errors use a single JSON envelope — see Errors.