Download docs/API_REFERENCE.md from shimen/yapapa: direct link, hf CLI and curl.
- Browser
- Download file 30.3 kB
-
https://huggingface.co/spaces/shimen/yapapa/resolve/main/docs/API_REFERENCE.md
- Command line
-
hf download hf://spaces/shimen/yapapa/docs/API_REFERENCE.md
-
curl -L -o API_REFERENCE.md https://huggingface.co/spaces/shimen/yapapa/resolve/main/docs/API_REFERENCE.md
wiwi β API Reference
All HTTP endpoints exposed by the wiwi gateway. Base URL is the wiwi process (default http://localhost:4000).
Auth summary:
- Client traffic (
/v1/*): virtual keysk-wiwi-β¦viaAuthorization: Bearer(orx-api-keyon/v1/messages). - Admin API (
/admin/*): master key viaAuthorization: Bearer. - User accounts (
/auth/*): signed HttpOnly session cookie. - Public (
/public/*): no auth.
1. Client surfaces (/v1/*)
POST /v1/chat/completions β OpenAI Chat dialect
Accepts the OpenAI Chat Completions request shape; returns OpenAI Chat responses (or SSE stream when "stream": true). Routing, failover, budgets, rate limits, and logging all apply.
- Auth:
Authorization: Bearer sk-wiwi-β¦ - Body: OpenAI Chat JSON (
model,messages,stream,tools,tool_choice,max_tokens/max_completion_tokens,temperature,top_p,stop,reasoning_effort,response_format, β¦) - Errors: OpenAI-shaped error body.
curl http://localhost:4000/v1/chat/completions \
-H "Authorization: Bearer sk-wiwi-β¦" -H "Content-Type: application/json" \
-d '{"model":"gpt-4o","messages":[{"role":"user","content":"hi"}],"stream":true}'
POST /v1/responses β OpenAI Responses dialect
OpenAI Responses API shape, used by Codex CLI and the OpenAI Agents SDK. Hosted tools (web_search, code_interpreter, β¦) translate through the IR builtin-tool registry to the backing provider.
- Auth:
Authorization: Bearer sk-wiwi-β¦ - Body: Responses JSON (
model,input,stream,instructions,tools,store,previous_response_id, β¦) - Events (streaming): Responses SSE event names (
response.created,response.output_text.delta,response.completed, β¦)
State (default on). A completed response is stored and may be continued by id: send previous_response_id with the next input instead of resending the transcript. The id in the response (resp_<request-id>) is wiwi's, not the upstream's, and reads/deletes are scoped to the presenting key β another key's id is a 404, exactly as if it never existed. store: false opts a single request out of persistence, and previous_response_id: <gone id> is a 404 not_found_error. Turn it off globally with wiwi_settings.store_responses: false.
GET /v1/responses/{id}β replay the stored response object (404 when absent, expired, or another key's).DELETE /v1/responses/{id}β{"id": β¦, "object": "response.deleted", "deleted": true}; 404 when nothing was removed.
POST /v1/messages β Anthropic Messages dialect
Anthropic Messages API shape, used by Claude Code and the Anthropic SDK. Also accepts x-api-key instead of Authorization.
- Auth:
Authorization: Bearer sk-wiwi-β¦orx-api-key: sk-wiwi-β¦+anthropic-versionheader - Body: Anthropic JSON (
model,messages,system,max_tokens,stream,tools,thinking, β¦) - Streaming events:
message_start,content_block_start/delta/stop,message_delta,message_stop(+ping,error).
A client that calls /v1/messages gets Anthropic-format responses even if the backing deployment is OpenAI β cross-dialect translation happens through the IR.
POST /v1/completions β legacy OpenAI Completions dialect
The original OpenAI text_completion shape: a bare prompt, no roles, no tool protocol. Used by older SDKs and by clients that predate chat/completions.
- Auth:
Authorization: Bearer sk-wiwi-β¦ - Body:
model,prompt(string),suffix,max_tokens/max_completion_tokens,temperature,top_p,stop,seed,n,stream,stream_options.include_usage - Returns:
{"id":"cmpl-<request-id>","object":"text_completion","model":β¦,"choices":[{"index":0,"text":β¦,"logprobs":null,"finish_reason":β¦}],"usage":{β¦}}; streaming emitstext_completionchunks thendata: [DONE] - Errors: OpenAI-shaped error body.
Refused with 400 invalid_request_error (no IR representation β refusing beats silently ignoring a parameter the caller sent): logprobs, best_of > 1, echo, n > 1, and prompt as a token-id array or a multi-element string array. prompt is mapped onto one user message; a prompt string plus suffix is concatenated.
POST /v1/messages/count_tokens
Anthropic token counting. Estimates tokens for the given Messages request (chars/4 heuristic when the provider doesn't expose a counting endpoint).
- Auth: same as
/v1/messages - Body: Anthropic Messages-shaped request
- Returns:
{"input_tokens": <int>}
GET /v1/models
OpenAI-style model list of the model groups visible to the authenticated key.
- Auth:
Authorization: Bearer sk-wiwi-β¦(or master key) - Returns:
{"object":"list","data":[{"id":"gpt-4o","object":"model",β¦}, β¦]}
Streaming headers & reconnect (all surfaces)
| Header | Meaning |
|---|---|
x-wiwi-request-id |
Request ID, echoed as a response header on every reply |
x-wiwi-cache |
HIT only when the reply was served from the gateway's exact-match response cache (see docs/CONFIG.md Β§ cache_settings). Absent on a miss. A hit still carries its own x-wiwi-request-id, and the body's id / created describe the request being answered, not the one that populated the entry |
x-wiwi-trace-id |
Trace ID (32 hex), only when telemetry.enabled. Pair it with the collector to pull the exact trace; see docs/CONFIG.md Β§ telemetry |
x-wiwi-stream-id |
Stream journal ID β pass back on reconnect to replay |
x-wiwi-translation-warnings |
Present only when the request asked for something the chosen provider could not honour β a provider-hosted tool the backend cannot run, a sampling parameter it has no field for, a dropped attachment, an unsigned thinking block. The messages are ;-separated. The same list is always in the request log (translation_warnings). Absent when nothing was dropped. |
x-wiwi-tool-args-warnings |
Present only when the model returned tool arguments that do not match the declared JSON Schema (wrong type, missing required property, unparseable JSON). The call is still forwarded as the model produced it β this is an advisory, not a rejection, because some clients repair the arguments themselves. ;-separated, same encoding as the translation warnings. Absent when every call validated. Unlike the translation warnings this list is not in the request log; the operator sees it as the tool_args_* proxy-log lines. |
x-wiwi-session-id |
Client session identity, used only when router_settings.session_affinity is on: pins the session to the deployment that served it so upstream prompt caches stay warm. A pin is dropped the moment that deployment is unhealthy or above this request's lane ceiling. Also accepted as a session_id query parameter for clients that cannot set headers; truncated to 128 chars. |
Last-Event-ID |
SSE standard header; reconnect replays missed frames. The value is the id: field of the last frame the client received β a monotonic per-frame sequence number, not the request id |
Every SSE frame carries an id: line holding a monotonic per-frame sequence number (stream_event_ids, on by default). That is the value a reconnecting client sends back as Last-Event-ID.
Stream journals are ON by default (.wiwi/journals/, 600 s TTL, 1 MiB cap): a client reconnecting with x-wiwi-stream-id + Last-Event-ID replays missed frames even after a wiwi restart. Mid-stream provider death resumes from the tape on a fallback deployment (continuation messages are synthesized from partial output).
WS /v1/realtime
Relays a Realtime WebSocket session to an OpenAI-shaped upstream. Off by
default (realtime.enabled); with it off the route does not exist.
ws://host/v1/realtime?model=<group>[&session_id=β¦]
Authorization: Bearer sk-wiwi-β¦
Every admission check runs before the upgrade, so a refusal is a readable
HTTP status rather than an opaque close: 401 bad key, 402 budget exhausted,
403 model not allowed, 404 unknown model or surface disabled, 429 rate
limited, 501 the provider has no realtime surface, 503 every deployment is
cooling or at its concurrency cap. Once the client sees 101, a session exists
upstream and errors arrive as close frames.
The refusal is a literal HTTP response written before the handshake completes,
not a close frame: uvicorn answers 403 to any close issued before accept
and discards the code, which would make a bad key indistinguishable from an
unknown model. Writing websocket.http.response.start by hand is what preserves
the status.
Frames pass through byte-for-byte in both directions β the session protocol is
stateful and ordered, so wiwi relays it and never rewrites it. A session holds
its deployment's concurrency slot for its whole life, so max_inflight,
priority lanes and session affinity apply to sessions exactly as they do to HTTP
requests.
Auth: header only by default. ?key= is accepted only when
realtime.allow_key_in_query is on β a query string lands in proxy logs,
browser history and access logs.
Billing: a session is charged once, at close, on the peak usage its events reported. Not the sum: the protocol restates overlapping windows of the same conversation, so summing inflates every multi-turn session. One request-log row is written per session.
2. Health & metrics
| Endpoint | Description |
|---|---|
GET /health |
Liveness. Always HTTP 200 when the process is up (the Docker healthcheck probes it). status is derived, not constant: ok when at least one provider is configured and at least one group has an available deployment, else degraded. Also reports groups, available_groups, providers, and the log-loss counters: dropped_request_logs (queue full), failed_request_log_writes (DB write failed), dropped_proxy_logs, failed_audit_log_writes, dropped_log_events (their sum), plus spend_charge_failures. A non-zero loss counter means durable accounting is missing rows. |
GET /metrics (default; path configurable) |
Prometheus text exposition: wiwi_requests_total, wiwi_request_duration_ms, wiwi_tokens_total{type=in|out|cached|reasoning}, wiwi_cost_total, wiwi_ttft_ms, wiwi_tps, wiwi_stream_errors_total, wiwi_prompt_cache_hits_total, wiwi_response_cache_hits_total, wiwi_request_logs_dropped_total, wiwi_spend_charge_failures_total. Process-lifetime counters (requests/tokens/cost/cache hits) are monotonic across scrapes; the windowed gauges and the three quantile families come from the in-memory LogEvent ring. wiwi_request_duration_ms/wiwi_ttft_ms/wiwi_tps are summary (percentiles computed at scrape time β no _bucket/_sum/_count series, so histogram_quantile() does not apply). wiwi_tps is output generation speed (completion tokens over the first-to-last-content-bearing-delta span, excluding queueing and prefill β text, thinking, a tool-call open, or a tool-call argument fragment, so tool-call-only streams are measured too and the window covers every token the numerator counts), so only streaming requests with a window over 0.05s contribute a sample; wiwi_tps_sample_ratio is a gauge for the share of the current window that wiwi_tps describes. |
3. Admin API (/admin/*, master-key auth)
Virtual keys
| Endpoint | Method | Description |
|---|---|---|
/admin/keys |
GET | List virtual keys (hashed at rest; never returns plaintext). |
/admin/keys/generate |
POST | Mint a key. Body: name, models, budget, rpm/tpm, expiry, priority (lane name). Plaintext sk-wiwi-β¦ returned only here. |
/admin/keys/{key_id} |
PATCH | Update name/models/budget/limits/lane. priority is operator-only: a key's owner cannot change their own lane. |
/admin/keys/{key_id} |
DELETE | Revoke a key. |
/admin/keys/{key_id}/disable |
POST | Disable without deleting. |
Providers & key pools
| Endpoint | Method | Description |
|---|---|---|
/admin/provider-catalog |
GET | Catalog of the 13 provider types (from PROVIDER_TYPES) with required fields. |
/admin/providers |
GET | List configured providers + key pools + health states. |
/admin/providers |
POST | Add a provider account. |
/admin/providers/{name} |
PATCH | Edit provider (name/type/base_url/timeout). |
/admin/providers/{name} |
DELETE | Delete provider + its keys, cascading away its deployments. |
/admin/providers/{name}/keys |
POST | Add a key to the provider's pool (label, secret, weight). |
/admin/providers/{name}/keys/{label} |
PATCH | Enable/disable, adjust weight. |
/admin/providers/{name}/keys/{label} |
DELETE | Remove one key from the pool. |
/admin/providers/{name}/keys/{label}/secret |
GET | Reveal key secret (admin action, audit-logged). |
/admin/providers/{name}/models |
GET | Models the provider exposes (live fetch where supported). |
/admin/providers/export |
GET | Export provider config (secrets masked or included per flag). |
/admin/providers/import |
POST | Import provider config. |
Backup and restore
Whole-gateway JSON, covering every table the console can mutate β not just
providers. The logic is wiwi/server/backup.py; the console surfaces it at
Settings β Import / Export.
| Endpoint | Method | Description |
|---|---|---|
/admin/backup |
GET | Export the whole gateway as one merge-restorable document. |
/admin/backup/import |
POST | Merge a document. Merge only β never deletes. |
GET /admin/backup query parameters (all optional):
| Param | Default | Description |
|---|---|---|
include_secrets |
off | Include plaintext provider key secrets and Cline refresh tokens. The only live credentials in the file. |
include_bodies |
off | Include request_body/response_body. A separate opt-in, so pulling credentials can never also pull user prompts. |
sections |
all non-bulk | Comma-separated subset of providers,keys,deployments,aliases,settings,prices,users,vkeys,logs,rollups,audit,responses. Bulk sections are excluded by default β the request log grows without limit. |
log_days |
30 |
Age bound on log sections. 0 = all time. |
log_limit |
0 |
Cap on raw log rows. 0 = no cap. A cap that bites is reported in truncated. |
POST /admin/backup/import takes the document verbatim. ?dry_run=1 runs the
full validation and returns the per-section plan without writing β the
console's Import panel shows that before enabling its Merge button. Any
validation error rejects the whole document with nothing applied.
Validation is not atomicity. The three bulk sections (logs, rollups,
responses) are applied first in one transaction, so a malformed row rolls
that whole pass back. The configuration sections then apply in order, each
persisting in its own transaction β so a failure partway leaves the earlier
config sections committed. The per-section counts in the response are the
report of what actually landed.
Merge keys, one per section: providers name Β· keys (provider_name, label) Β·
deployments (group_name, provider_name, model_id) Β· aliases the alias name Β·
settings the setting key Β· prices model_id (+ (model_id, scope) per scoped
override) Β· users username Β· virtual keys key_hash Β· request logs
request_id Β· rollups (bucket_ts, key_id, model_group, provider, serving_model).
Four things a backup deliberately does not do:
- It does not reveal virtual keys or passwords. Only their hashes are
stored (
hash_keyis unsalted SHA-256), so the file carries digests. Importing a digest does restore a working key β the same plaintext authenticates on the target β but no key can be printed again afterwards. - It never replays the audit log.
audit_logshas no natural key, and restoring another instance's mutation trail would forge this one's operator history. The rows are exported and reported asaudit_skipped. - It does not carry the master key. That lives in
wiwi.yamland the environment; a restored instance needsWIWI_MASTER_KEYset separately or it fails closed at startup. - It does not delete. A deployment or alias the source recorded as removed is removed here too, but nothing is ever dropped merely for being absent from the file.
OAuth providers (Cline, WorkBuddy)
| Endpoint | Method | Description |
|---|---|---|
/admin/cline/models |
GET | Models available to the connected Cline account. |
/admin/cline/settings |
GET / PUT | Cline account settings. |
/admin/cline/settings/default-models/{model_id} |
DELETE | Remove a default-model mapping. |
/admin/cline/oauth/login-url |
POST | Get OAuth login URL. |
/admin/cline/oauth/connect |
POST | Complete OAuth connect. |
/admin/cline/oauth/auto-connect |
POST | Initiate the redirect-based connect. Dormant: Cline's Google OAuth ignores callback_url, so the redirect never returns a ?code=; both Cline UIs use the paste-code flow (login-url + connect) instead. The route works and is tested; it becomes usable if Cline honours callback_url. |
/admin/cline/oauth/status |
GET | Connection/refresh state. |
/admin/cline/oauth/refresh |
POST | Force token refresh. |
/admin/cline/oauth/disconnect |
DELETE | Disconnect account. |
/cline/oauth/callback |
GET | OAuth redirect target (browser). |
OAuth providers (Hermes / Nous Portal β device-code flow)
Hermes uses OAuth 2.0 device authorization (RFC 8628). Unlike Cline's paste-code flow there is no callback URL: wiwi mints a device code and verification URL, the operator approves in a browser, and wiwi polls the token endpoint. On approval the account is appended as its own pool key holding a hermes:<json> credential blob (the adapter resolves the Authorization: Bearer from it), so multiple Nous accounts share one provider's key pool and round-robin β each account has its own rotating refresh token and its own rate-limit bucket. No API key is required β the device-code call bootstraps the account (creating it with a placeholder key if it does not exist). Tokens are rotated automatically by a background sweeper (~2 min before expiry) and on a mid-session 401; the refresh endpoints below force one.
| Endpoint | Method | Description |
|---|---|---|
/admin/hermes/oauth/device-code?provider=<name> |
POST | Start device authorization. Returns {state, user_code, verification_uri, verification_uri_complete, expires_in, interval}. Creates the account if it does not exist. |
/admin/hermes/oauth/poll |
POST | One poll attempt for a pending session ({"state": β¦}). Returns {"pending": true} while unapproved, or the connected account ({"connected": true, "key_label", "access_token_masked", "expires_at"}) on approval β the account is added as a new pool key. 410 when the session's 10-minute TTL lapses; 404 for an unknown state. |
/admin/hermes/oauth/status?provider=<name> |
GET | {connected, accounts[]} β one entry per connected account with {key_label, managed, label, expires_at, needs_refresh, has_refresh_token}. |
/admin/hermes/oauth/refresh |
POST | Force a token refresh ({"provider": β¦, "key_label"?: β¦}). key_label targets one account (default: the first managed one). 401 on a terminal grant (invalid_grant/reused refresh token) β that account needs re-login; others keep serving. |
/admin/hermes/oauth/disconnect |
DELETE | Remove accounts ({"provider": β¦, "key_label"?: β¦}). With key_label removes just that account; without it, removes every managed account. Returns {disconnected: <count>}. |
| /admin/workbuddy/accounts | GET | List connected WorkBuddy accounts. |
| /admin/workbuddy/import | POST | Import WorkBuddy account credentials. |
| /admin/workbuddy/export | GET | Export account data. |
| /admin/workbuddy/refresh | POST | Force token refresh. |
Model groups, deployments, aliases
| Endpoint | Method | Description |
|---|---|---|
/admin/models |
GET | All model groups + deployments. |
/admin/model-groups/{name}/deployments |
POST | Add a deployment to a group. |
/admin/model-groups/{name}/deployments |
DELETE | Remove a deployment. |
/admin/model-groups/{name} |
PATCH | Edit group (routing/alias settings). |
/admin/aliases |
POST | Create a model alias β group mapping. |
Pricing & alerts
| Endpoint | Method | Description |
|---|---|---|
/admin/pricing |
GET | Per-model token prices (DB-backed model_prices), each with its per-provider scopes. |
/admin/pricing/{model_id} |
PUT | Set/override price for a model. Add ?provider=<account-or-type> to scope it to one provider; omit for the all-providers base rate. |
/admin/pricing/{model_id} |
DELETE | Remove override (fall back to bundled table). With ?provider= removes only that scope. |
/admin/alert-rules |
GET / PUT | Spend/alert rule configuration. |
Logs, stats, realtime
| Endpoint | Method | Description |
|---|---|---|
/admin/logs/requests |
GET | Request logs (paginated; supports key/model/status/time filters). |
/admin/logs/proxy |
GET | Proxy stream log (upstream request/response metadata). |
/admin/stats/overview |
GET | Rollup: requests/min, token totals, cache-hit %, avg TPS, p95 TTFT, error rate, spend. Ranges β€ 24 h read the in-memory ring; 7d/30d/all-time read DB aggregates. |
/admin/stats/timeseries |
GET | Bucketed token/tps series (bucket size scales with range: 1 min β 1 day). |
/admin/stream |
GET (SSE) | Realtime event stream consumed by the UI (live stats, log events). |
Users
| Endpoint | Method | Description |
|---|---|---|
/admin/users |
GET | List user accounts. |
/admin/users/{uid} |
PATCH | Change role, disable, reset. |
4. User accounts (/auth/*)
Cookie-session auth for the multi-user UI (separate from virtual keys). Passwords hashed with stdlib PBKDF2; sessions are signed HMAC cookies.
| Endpoint | Method | Description |
|---|---|---|
/auth/signup |
POST | Create an account (username, password). |
/auth/login |
POST | Log in; sets the session cookie. |
/auth/logout |
POST | Clear session. |
/auth/me |
GET | Current user + role. |
/auth/playground-key |
POST | Mint a scoped temporary virtual key for the playground. |
Roles: normal users see only their own keys' usage in /app/*; admins see the full console.
5. Public (/public/*, no auth)
| Endpoint | Method | Description |
|---|---|---|
/public/models |
GET | Secret-free model catalog for the public front (names + metadata, no keys, no prices beyond published ones). |
6. Error bodies
WiwiError is the unified internal error type; each wire dialect owns its error_body() mapping so clients always see errors in their own dialect:
| Surface | Shape |
|---|---|
| OpenAI Chat | {"error": {"message", "type", "code"}} |
| OpenAI Completions | {"error": {"message", "type", "code"}} (same envelope as Chat) |
| OpenAI Responses | Responses-style error event / object |
| Anthropic Messages | {"type":"error","error":{"type":"<anthropic_error_type>","message":β¦}} |
Common status codes: 401 (bad key), 402 (budget cap exceeded), 404 (unknown model/group), 413 (body too large), 429 (rate limit β a per-key or per-deployment quota with a horizon), 502/503 (upstream unavailable after retries/failover), 503 + Retry-After (shed: every deployment for the group is at its concurrency ceiling for this request's lane β saturated now, not quota spent).
7. Worked examples
The tables above are the contract; these are copy-pasteable equivalents.
MK="Authorization: Bearer $WIWI_MASTER_KEY"
Virtual keys
# mint β budget / RPM / TPM / model allowlist / TTL / optional custom_key
curl -X POST localhost:4000/admin/keys/generate -H "$MK" \
-d '{"name": "team-a", "max_budget": 10, "rpm": 60, "tpm": 100000,
"models": ["gpt-4o"], "ttl_seconds": 86400}'
# β {"key":"sk-wiwi-...","id":"k...","note":"store this key now..."}
# mint into a priority lane β must name a lane in router_settings.priority_lanes,
# or the call is a 400 rather than a silent fall back to default_lane
curl -X POST localhost:4000/admin/keys/generate -H "$MK" \
-d '{"name": "nightly", "priority": "bulk"}'
curl localhost:4000/admin/keys -H "$MK"
curl -X PATCH localhost:4000/admin/keys/<id> -H "$MK" -d '{"max_budget": 20}'
curl -X POST localhost:4000/admin/keys/<id>/disable -H "$MK"
curl -X DELETE localhost:4000/admin/keys/<id> -H "$MK"
Providers & key pools
curl localhost:4000/admin/provider-catalog -H "$MK" # 11 built-in cards + configured?
curl localhost:4000/admin/providers -H "$MK" # pool status: health + cooldowns
curl -X POST localhost:4000/admin/providers -H "$MK" \
-d '{"name": "openai-backup", "provider_type": "openai",
"base_url": "https://api.openai.com/v1", "key": "os.environ/BACKUP_KEY"}'
curl -X PATCH localhost:4000/admin/providers/<name> -H "$MK" -d '{"name": "openai-primary"}'
curl -X DELETE localhost:4000/admin/providers/<name> -H "$MK" # 409 while groups reference it
# key pool
curl -X POST localhost:4000/admin/providers/<name>/keys -H "$MK" \
-d '{"label": "extra", "key": "os.environ/EXTRA_KEY", "weight": 2}'
curl -X PATCH localhost:4000/admin/providers/<name>/keys/<label> -H "$MK" \
-d '{"disabled": true, "weight": 5}' # + reset_status: true clears cooldown
curl -X DELETE localhost:4000/admin/providers/<name>/keys/<label> -H "$MK"
curl localhost:4000/admin/providers/<name>/keys/<label>/secret -H "$MK" # audit-logged reveal
curl localhost:4000/admin/providers/<name>/models -H "$MK" # live upstream model ids
Models, groups, aliases
curl localhost:4000/admin/models -H "$MK"
curl -X PATCH localhost:4000/admin/model-groups/<name> -H "$MK" \
-d '{"weights": {"openai-main/gpt-4o": 3}, "strategy": "least-busy"}'
curl -X POST localhost:4000/admin/model-groups/<name>/deployments -H "$MK" \
-d '{"group": "gpt-4o", "provider": "openrouter", "model_id": "openai/gpt-4o", "weight": 1}'
curl -X DELETE localhost:4000/admin/model-groups/<name>/deployments -H "$MK" -d '{...}'
curl -X POST localhost:4000/admin/aliases -H "$MK" \
-d '{"set": {"gpt-4": "gpt-4o"}, "unset": ["gpt-3.5"]}'
Pricing, logs, stats, users
curl localhost:4000/admin/pricing -H "$MK"
curl -X PUT localhost:4000/admin/pricing/<model_id> -H "$MK" -d '{...}'
# Per-provider prices: add ?provider=<account-or-type>. Omit it to set the base rate.
curl -X PUT "localhost:4000/admin/pricing/<model_id>?provider=openai-main" -H "$MK" \
-d '{"input_per_1m": 1.0, "output_per_1m": 2.0}'
curl localhost:4000/admin/logs/requests -H "$MK" # DB-backed
curl localhost:4000/admin/logs/proxy -H "$MK" # ring buffer
curl localhost:4000/admin/stats/overview -H "$MK" # p50/p95/p99, cost, tokens
curl "localhost:4000/admin/stats/timeseries?bucket=minute&metric=cost&minutes=60" -H "$MK"
curl localhost:4000/admin/stream -H "$MK" # SSE live tail
curl localhost:4000/admin/alert-rules -H "$MK"
curl -X PUT localhost:4000/admin/alert-rules -H "$MK" -d '{...}'
curl localhost:4000/admin/users -H "$MK"
curl -X PATCH localhost:4000/admin/users/<uid> -H "$MK" -d '{"role": "admin"}'
Session auth
curl -X POST localhost:4000/auth/signup -d '{"email": "...", "password": "..."}'
curl -X POST localhost:4000/auth/login -d '{"email": "...", "password": "..."}'
curl -X POST localhost:4000/auth/logout
curl localhost:4000/auth/me
curl -X POST localhost:4000/auth/playground-key # scoped 24h session key
OAuth providers (Cline / WorkBuddy)
# Cline
curl -X POST localhost:4000/admin/cline/oauth/login-url -H "$MK" -d '{}' # {auth_url, state}
curl -X POST localhost:4000/admin/cline/oauth/connect -H "$MK" -d '{"code": "..."}'
curl -X POST localhost:4000/admin/cline/oauth/auto-connect -H "$MK" -d '{}'
curl localhost:4000/admin/cline/oauth/status -H "$MK"
curl -X POST localhost:4000/admin/cline/oauth/refresh -H "$MK" -d '{"provider": "cline-main"}'
curl -X DELETE localhost:4000/admin/cline/oauth/disconnect -H "$MK" -d '{"provider": "cline-main"}'
# Cline global model list β pick ids once, auto-deploy to every Cline account
curl localhost:4000/admin/cline/models -H "$MK"
curl -X PUT localhost:4000/admin/cline/settings -H "$MK" \
-d '{"default_models": ["anthropic/claude-sonnet-4-5", "openai/gpt-4o"]}'
# Hermes (Nous Portal) β device-code OAuth. No API key needed: the first call
# bootstraps the account. Open verification_uri_complete in a browser, approve,
# then poll until it returns connected:true. Each approval ADDS an account as a
# new pool key (key_label in the response); repeat to connect more accounts and
# they round-robin. Tokens auto-refresh ~2 min before expiry and on a 401.
curl -X POST "localhost:4000/admin/hermes/oauth/device-code?provider=hermes-main" -H "$MK"
# -> {state, user_code, verification_uri_complete, expires_in, interval}
curl -X POST localhost:4000/admin/hermes/oauth/poll -H "$MK" -d '{"state": "..."}'
# -> {"pending": true} until approved, then {"connected": true, "key_label": ...}
curl localhost:4000/admin/hermes/oauth/status?provider=hermes-main -H "$MK"
# -> {"connected": true, "accounts": [{"key_label", "expires_at", ...}, ...]}
# Force a rotation for one account (key_label optional; defaults to the first):
curl -X POST localhost:4000/admin/hermes/oauth/refresh -H "$MK" -d '{"provider": "hermes-main", "key_label": "acct@example.com"}'
# Disconnect one account, or the whole pool when key_label is omitted:
curl -X DELETE localhost:4000/admin/hermes/oauth/disconnect -H "$MK" -d '{"provider": "hermes-main", "key_label": "acct@example.com"}'
# WorkBuddy (CodeBuddy) β parallel API
curl localhost:4000/admin/workbuddy/accounts -H "$MK"
curl -X POST localhost:4000/admin/workbuddy/import -H "$MK" -d '{"accounts": [...]}'
curl -X POST localhost:4000/admin/workbuddy/export -H "$MK"
curl -X POST localhost:4000/admin/workbuddy/refresh -H "$MK" -d '{"label": "main"}'