tokenbiryani 0.2.0__py3-none-any.whl

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (51) hide show
  1. tokenbiryani/__init__.py +3 -0
  2. tokenbiryani/api/__init__.py +0 -0
  3. tokenbiryani/api/app.py +583 -0
  4. tokenbiryani/api/asgi.py +32 -0
  5. tokenbiryani/cli.py +1045 -0
  6. tokenbiryani/config.py +532 -0
  7. tokenbiryani/core/__init__.py +0 -0
  8. tokenbiryani/core/account.py +258 -0
  9. tokenbiryani/core/batch.py +135 -0
  10. tokenbiryani/core/breaker.py +53 -0
  11. tokenbiryani/core/cacheadvice.py +239 -0
  12. tokenbiryani/core/diagnostics.py +131 -0
  13. tokenbiryani/core/estimator.py +180 -0
  14. tokenbiryani/core/gateway.py +2395 -0
  15. tokenbiryani/core/handoff.py +87 -0
  16. tokenbiryani/core/keys.py +199 -0
  17. tokenbiryani/core/limits.py +440 -0
  18. tokenbiryani/core/oauth.py +222 -0
  19. tokenbiryani/core/pacing.py +320 -0
  20. tokenbiryani/core/queue.py +132 -0
  21. tokenbiryani/core/router.py +323 -0
  22. tokenbiryani/core/secrets.py +114 -0
  23. tokenbiryani/core/session.py +117 -0
  24. tokenbiryani/dashboard/__init__.py +56 -0
  25. tokenbiryani/dashboard/console.css +610 -0
  26. tokenbiryani/dashboard/console.html +3250 -0
  27. tokenbiryani/observability/__init__.py +0 -0
  28. tokenbiryani/observability/events.py +171 -0
  29. tokenbiryani/observability/usage.py +226 -0
  30. tokenbiryani/prices.yaml +77 -0
  31. tokenbiryani/providers/__init__.py +0 -0
  32. tokenbiryani/providers/anthropic_api.py +118 -0
  33. tokenbiryani/providers/base.py +173 -0
  34. tokenbiryani/providers/bedrock.py +182 -0
  35. tokenbiryani/providers/oauth.py +165 -0
  36. tokenbiryani/providers/oauth_credentials.py +293 -0
  37. tokenbiryani/providers/translate.py +35 -0
  38. tokenbiryani/providers/vertex.py +144 -0
  39. tokenbiryani/proxy/__init__.py +0 -0
  40. tokenbiryani/proxy/errors.py +169 -0
  41. tokenbiryani/proxy/sse.py +98 -0
  42. tokenbiryani/store/__init__.py +0 -0
  43. tokenbiryani/store/base.py +150 -0
  44. tokenbiryani/store/memory.py +149 -0
  45. tokenbiryani/store/redis_store.py +222 -0
  46. tokenbiryani/store/sqlite.py +336 -0
  47. tokenbiryani-0.2.0.dist-info/METADATA +697 -0
  48. tokenbiryani-0.2.0.dist-info/RECORD +51 -0
  49. tokenbiryani-0.2.0.dist-info/WHEEL +4 -0
  50. tokenbiryani-0.2.0.dist-info/entry_points.txt +2 -0
  51. tokenbiryani-0.2.0.dist-info/licenses/LICENSE +202 -0
@@ -0,0 +1,697 @@
1
+ Metadata-Version: 2.5
2
+ Name: tokenbiryani
3
+ Version: 0.2.0
4
+ Summary: A pooling gateway for Claude accounts: routes on live rate-limit headers, keeps prompt caches warm, fails over without dropping a stream.
5
+ Project-URL: Homepage, https://github.com/Abhi-shekes/tokenbiryani
6
+ Project-URL: Documentation, https://abhi-shekes.github.io/tokenbiryani/
7
+ Project-URL: Repository, https://github.com/Abhi-shekes/tokenbiryani
8
+ Project-URL: Issues, https://github.com/Abhi-shekes/tokenbiryani/issues
9
+ Project-URL: Changelog, https://github.com/Abhi-shekes/tokenbiryani/blob/main/CHANGELOG.md
10
+ License:
11
+ Apache License
12
+ Version 2.0, January 2004
13
+ http://www.apache.org/licenses/
14
+
15
+ TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
16
+
17
+ 1. Definitions.
18
+
19
+ "License" shall mean the terms and conditions for use, reproduction,
20
+ and distribution as defined by Sections 1 through 9 of this document.
21
+
22
+ "Licensor" shall mean the copyright owner or entity authorized by
23
+ the copyright owner that is granting the License.
24
+
25
+ "Legal Entity" shall mean the union of the acting entity and all
26
+ other entities that control, are controlled by, or are under common
27
+ control with that entity. For the purposes of this definition,
28
+ "control" means (i) the power, direct or indirect, to cause the
29
+ direction or management of such entity, whether by contract or
30
+ otherwise, or (ii) ownership of fifty percent (50%) or more of the
31
+ outstanding shares, or (iii) beneficial ownership of such entity.
32
+
33
+ "You" (or "Your") shall mean an individual or Legal Entity
34
+ exercising permissions granted by this License.
35
+
36
+ "Source" form shall mean the preferred form for making modifications,
37
+ including but not limited to software source code, documentation
38
+ source, and configuration files.
39
+
40
+ "Object" form shall mean any form resulting from mechanical
41
+ transformation or translation of a Source form, including but
42
+ not limited to compiled object code, generated documentation,
43
+ and conversions to other media types.
44
+
45
+ "Work" shall mean the work of authorship, whether in Source or
46
+ Object form, made available under the License, as indicated by a
47
+ copyright notice that is included in or attached to the work
48
+ (an example is provided in the Appendix below).
49
+
50
+ "Derivative Works" shall mean any work, whether in Source or Object
51
+ form, that is based on (or derived from) the Work and for which the
52
+ editorial revisions, annotations, elaborations, or other modifications
53
+ represent, as a whole, an original work of authorship. For the purposes
54
+ of this License, Derivative Works shall not include works that remain
55
+ separable from, or merely link (or bind by name) to the interfaces of,
56
+ the Work and Derivative Works thereof.
57
+
58
+ "Contribution" shall mean any work of authorship, including
59
+ the original version of the Work and any modifications or additions
60
+ to that Work or Derivative Works thereof, that is intentionally
61
+ submitted to Licensor for inclusion in the Work by the copyright owner
62
+ or by an individual or Legal Entity authorized to submit on behalf of
63
+ the copyright owner. For the purposes of this definition, "submitted"
64
+ means any form of electronic, verbal, or written communication sent
65
+ to the Licensor or its representatives, including but not limited to
66
+ communication on electronic mailing lists, source code control systems,
67
+ and issue tracking systems that are managed by, or on behalf of, the
68
+ Licensor for the purpose of discussing and improving the Work, but
69
+ excluding communication that is conspicuously marked or otherwise
70
+ designated in writing by the copyright owner as "Not a Contribution."
71
+
72
+ "Contributor" shall mean Licensor and any individual or Legal Entity
73
+ on behalf of whom a Contribution has been received by Licensor and
74
+ subsequently incorporated within the Work.
75
+
76
+ 2. Grant of Copyright License. Subject to the terms and conditions of
77
+ this License, each Contributor hereby grants to You a perpetual,
78
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
79
+ copyright license to reproduce, prepare Derivative Works of,
80
+ publicly display, publicly perform, sublicense, and distribute the
81
+ Work and such Derivative Works in Source or Object form.
82
+
83
+ 3. Grant of Patent License. Subject to the terms and conditions of
84
+ this License, each Contributor hereby grants to You a perpetual,
85
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
86
+ (except as stated in this section) patent license to make, have made,
87
+ use, offer to sell, sell, import, and otherwise transfer the Work,
88
+ where such license applies only to those patent claims licensable
89
+ by such Contributor that are necessarily infringed by their
90
+ Contribution(s) alone or by combination of their Contribution(s)
91
+ with the Work to which such Contribution(s) was submitted. If You
92
+ institute patent litigation against any entity (including a
93
+ cross-claim or counterclaim in a lawsuit) alleging that the Work
94
+ or a Contribution incorporated within the Work constitutes direct
95
+ or contributory patent infringement, then any patent licenses
96
+ granted to You under this License for that Work shall terminate
97
+ as of the date such litigation is filed.
98
+
99
+ 4. Redistribution. You may reproduce and distribute copies of the
100
+ Work or Derivative Works thereof in any medium, with or without
101
+ modifications, and in Source or Object form, provided that You
102
+ meet the following conditions:
103
+
104
+ (a) You must give any other recipients of the Work or
105
+ Derivative Works a copy of this License; and
106
+
107
+ (b) You must cause any modified files to carry prominent notices
108
+ stating that You changed the files; and
109
+
110
+ (c) You must retain, in the Source form of any Derivative Works
111
+ that You distribute, all copyright, patent, trademark, and
112
+ attribution notices from the Source form of the Work,
113
+ excluding those notices that do not pertain to any part of
114
+ the Derivative Works; and
115
+
116
+ (d) If the Work includes a "NOTICE" text file as part of its
117
+ distribution, then any Derivative Works that You distribute must
118
+ include a readable copy of the attribution notices contained
119
+ within such NOTICE file, excluding those notices that do not
120
+ pertain to any part of the Derivative Works, in at least one
121
+ of the following places: within a NOTICE text file distributed
122
+ as part of the Derivative Works; within the Source form or
123
+ documentation, if provided along with the Derivative Works; or,
124
+ within a display generated by the Derivative Works, if and
125
+ wherever such third-party notices normally appear. The contents
126
+ of the NOTICE file are for informational purposes only and
127
+ do not modify the License. You may add Your own attribution
128
+ notices within Derivative Works that You distribute, alongside
129
+ or as an addendum to the NOTICE text from the Work, provided
130
+ that such additional attribution notices cannot be construed
131
+ as modifying the License.
132
+
133
+ You may add Your own copyright statement to Your modifications and
134
+ may provide additional or different license terms and conditions
135
+ for use, reproduction, or distribution of Your modifications, or
136
+ for any such Derivative Works as a whole, provided Your use,
137
+ reproduction, and distribution of the Work otherwise complies with
138
+ the conditions stated in this License.
139
+
140
+ 5. Submission of Contributions. Unless You explicitly state otherwise,
141
+ any Contribution intentionally submitted for inclusion in the Work
142
+ by You to the Licensor shall be under the terms and conditions of
143
+ this License, without any additional terms or conditions.
144
+ Notwithstanding the above, nothing herein shall supersede or modify
145
+ the terms of any separate license agreement you may have executed
146
+ with Licensor regarding such Contributions.
147
+
148
+ 6. Trademarks. This License does not grant permission to use the trade
149
+ names, trademarks, service marks, or product names of the Licensor,
150
+ except as required for reasonable and customary use in describing the
151
+ origin of the Work and reproducing the content of the NOTICE file.
152
+
153
+ 7. Disclaimer of Warranty. Unless required by applicable law or
154
+ agreed to in writing, Licensor provides the Work (and each
155
+ Contributor provides its Contributions) on an "AS IS" BASIS,
156
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
157
+ implied, including, without limitation, any warranties or conditions
158
+ of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
159
+ PARTICULAR PURPOSE. You are solely responsible for determining the
160
+ appropriateness of using or redistributing the Work and assume any
161
+ risks associated with Your exercise of permissions under this License.
162
+
163
+ 8. Limitation of Liability. In no event and under no legal theory,
164
+ whether in tort (including negligence), contract, or otherwise,
165
+ unless required by applicable law (such as deliberate and grossly
166
+ negligent acts) or agreed to in writing, shall any Contributor be
167
+ liable to You for damages, including any direct, indirect, special,
168
+ incidental, or consequential damages of any character arising as a
169
+ result of this License or out of the use or inability to use the
170
+ Work (including but not limited to damages for loss of goodwill,
171
+ work stoppage, computer failure or malfunction, or any and all
172
+ other commercial damages or losses), even if such Contributor
173
+ has been advised of the possibility of such damages.
174
+
175
+ 9. Accepting Warranty or Additional Liability. While redistributing
176
+ the Work or Derivative Works thereof, You may choose to offer,
177
+ and charge a fee for, acceptance of support, warranty, indemnity,
178
+ or other liability obligations and/or rights consistent with this
179
+ License. However, in accepting such obligations, You may act only
180
+ on Your own behalf and on Your sole responsibility, not on behalf
181
+ of any other Contributor, and only if You agree to indemnify,
182
+ defend, and hold each Contributor harmless for any liability
183
+ incurred by, or claims asserted against, such Contributor by reason
184
+ of your accepting any such warranty or additional liability.
185
+
186
+ END OF TERMS AND CONDITIONS
187
+
188
+ APPENDIX: How to apply the Apache License to your work.
189
+
190
+ To apply the Apache License to your work, attach the following
191
+ boilerplate notice, with the fields enclosed by brackets "[]"
192
+ replaced with your own identifying information. (Don't include
193
+ the brackets!) The text should be enclosed in the appropriate
194
+ comment syntax for the file format. We also recommend that a
195
+ file or class name and description of purpose be included on the
196
+ same "printed page" as the copyright notice for easier
197
+ identification within third-party archives.
198
+
199
+ Copyright [yyyy] [name of copyright owner]
200
+
201
+ Licensed under the Apache License, Version 2.0 (the "License");
202
+ you may not use this file except in compliance with the License.
203
+ You may obtain a copy of the License at
204
+
205
+ http://www.apache.org/licenses/LICENSE-2.0
206
+
207
+ Unless required by applicable law or agreed to in writing, software
208
+ distributed under the License is distributed on an "AS IS" BASIS,
209
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
210
+ See the License for the specific language governing permissions and
211
+ limitations under the License.
212
+ License-File: LICENSE
213
+ Keywords: anthropic,claude,claude-code,gateway,llm-gateway,load-balancing,prompt-caching,proxy,rate-limit
214
+ Classifier: Development Status :: 3 - Alpha
215
+ Classifier: Intended Audience :: Developers
216
+ Classifier: License :: OSI Approved :: Apache Software License
217
+ Classifier: Programming Language :: Python :: 3
218
+ Classifier: Topic :: Internet :: Proxy Servers
219
+ Requires-Python: >=3.8
220
+ Requires-Dist: cryptography>=41.0
221
+ Requires-Dist: fastapi>=0.100
222
+ Requires-Dist: httpx>=0.24
223
+ Requires-Dist: pyyaml>=6.0
224
+ Requires-Dist: uvicorn>=0.23
225
+ Provides-Extra: all
226
+ Requires-Dist: botocore>=1.31; extra == 'all'
227
+ Requires-Dist: google-auth>=2.20; extra == 'all'
228
+ Requires-Dist: redis>=5.0; extra == 'all'
229
+ Provides-Extra: bedrock
230
+ Requires-Dist: botocore>=1.31; extra == 'bedrock'
231
+ Provides-Extra: dev
232
+ Requires-Dist: botocore>=1.31; extra == 'dev'
233
+ Requires-Dist: fakeredis>=2.20; extra == 'dev'
234
+ Requires-Dist: google-auth>=2.20; extra == 'dev'
235
+ Requires-Dist: mypy>=1.8; extra == 'dev'
236
+ Requires-Dist: playwright>=1.40; extra == 'dev'
237
+ Requires-Dist: pytest-asyncio>=0.21; extra == 'dev'
238
+ Requires-Dist: pytest-cov>=4.1; extra == 'dev'
239
+ Requires-Dist: pytest>=7.4; extra == 'dev'
240
+ Requires-Dist: redis>=5.0; extra == 'dev'
241
+ Requires-Dist: ruff>=0.1; extra == 'dev'
242
+ Requires-Dist: types-pyyaml; extra == 'dev'
243
+ Requires-Dist: watchfiles>=0.21; extra == 'dev'
244
+ Provides-Extra: docs
245
+ Requires-Dist: mkdocs-material<10,>=9.5; extra == 'docs'
246
+ Provides-Extra: redis
247
+ Requires-Dist: redis>=5.0; extra == 'redis'
248
+ Provides-Extra: secrets
249
+ Provides-Extra: vertex
250
+ Requires-Dist: google-auth>=2.20; extra == 'vertex'
251
+ Description-Content-Type: text/markdown
252
+
253
+ # tokenbiryani
254
+
255
+ **A pooling gateway for Claude accounts.** One Anthropic-compatible endpoint in front of
256
+ every credential you own. It routes on live rate-limit headers, keeps a conversation on
257
+ the account holding its prompt cache, fails over without dropping a stream, and queues
258
+ honestly when the whole pool is dry.
259
+
260
+ ```bash
261
+ pip install tokenbiryani # or: pipx install tokenbiryani
262
+ tokenbiryani init # writes tokenbiryani.yaml + a virtual key
263
+ tokenbiryani serve # starts with an empty pool
264
+ tokenbiryani console # opens the browser, already signed in
265
+ ```
266
+
267
+ ![The console's Overview screen: readiness, next reset, queue depth, cache hit rate and
268
+ spend across the top, a capacity horizon below it, and every account in the pool with
269
+ its headroom meters, p95, cache rate and spend](docs/images/console-overview.png)
270
+
271
+ Nothing needs exporting first. The console's wizard takes your first credential,
272
+ **verifies it** — the same rate-limit-header check `tokenbiryani doctor` performs — and
273
+ hands you the two lines that use it, with a working key already in them:
274
+
275
+ ```bash
276
+ export ANTHROPIC_BASE_URL=http://localhost:8787
277
+ export ANTHROPIC_AUTH_TOKEN=bir_...
278
+ claude # Claude Code now runs through the pool
279
+ ```
280
+
281
+ Prefer the terminal:
282
+
283
+ ```bash
284
+ tokenbiryani accounts add work --api-key sk-ant-... # probed before it is stored
285
+ tokenbiryani accounts test # all of them, headers included
286
+ ```
287
+
288
+ No client changes. The gateway speaks the Messages API verbatim — it swaps the auth
289
+ header and picks an upstream, and touches nothing else in the request.
290
+
291
+ > The one exception is the Bedrock and Vertex adapters. Those platforms address the
292
+ > model in the URL and stamp their own `anthropic_version`, so exactly two fields are
293
+ > translated, in one file (`providers/translate.py`), and nowhere else. Everything else
294
+ > the caller sent — including parameters this gateway has never heard of — travels
295
+ > through untouched.
296
+
297
+ ---
298
+
299
+ ## Why not just use a generic proxy
300
+
301
+ LiteLLM, Portkey and friends are breadth plays: many providers, lowest-common-denominator
302
+ routing. This is a depth play on one provider, and four things fall out of that.
303
+
304
+ | | tokenbiryani | Generic proxy |
305
+ |---|---|---|
306
+ | **Routing signal** | Live mirror of `anthropic-ratelimit-*` headers — real remaining request and token budget, per account, per reset window | Round-robin, or reactive backoff after a 429 lands |
307
+ | **Prompt cache** | Session affinity keeps a conversation on the account holding its cache; cache-break is a tracked metric | Round-robin shreds the cache silently |
308
+ | **Claude Code** | First-class client — long streaming turns, huge cached prefixes, tool loops | Treated as generic chat completion |
309
+ | **Failover** | Transparent up to the first streamed token, with an explicit documented boundary | Usually all-or-nothing |
310
+
311
+ ### The benchmark
312
+
313
+ The same workload — 24 concurrent conversations, 8 turns each, across 4 accounts —
314
+ under each strategy, against a mock upstream that models Anthropic's per-credential
315
+ prompt cache. Prices are illustrative ratios, not a price list.
316
+
317
+ | Strategy | Cache hit | Cache breaks | Cost | vs sticky | Billed input |
318
+ |---|---|---|---|---|---|
319
+ | sticky_headroom | 79.6% | 0 | $0.4774 | — | 433,152 |
320
+ | round_robin | 47.8% | 144 | $0.9535 | 2.00x | 433,152 |
321
+ | least_loaded | 47.8% | 144 | $0.9535 | 2.00x | 433,152 |
322
+ | headroom | 47.8% | 144 | $0.9535 | 2.00x | 433,152 |
323
+
324
+ **Cache-blind routing costs 2.00x here.** And note that round-robin, least-loaded and
325
+ most-headroom all pay exactly the same penalty: any strategy that ignores affinity
326
+ visits every account once per conversation, so they all take the same number of cache
327
+ misses. The penalty is inherent to cache-blindness, not a quirk of round-robin.
328
+
329
+ Reproduce it with `python benchmarks/cache_affinity.py`. `tests/test_benchmark.py` fails if sticky ever stops winning.
330
+
331
+ ### The prompt cache is the constraint
332
+
333
+ Anthropic's cache is scoped per credential. A Claude Code turn resends a large stable
334
+ prefix each time; warm, it bills at roughly a tenth of the input rate. **A gateway that
335
+ balances load without cache awareness can cost several times more than no gateway at
336
+ all.** So the default strategy is sticky, not balanced: affinity first, rebalance only
337
+ when the owner genuinely cannot serve. Every forced break is counted.
338
+
339
+ ---
340
+
341
+ ## How a request flows
342
+
343
+ ```
344
+ auth → admit → route → lease → proxy → reconcile → recover
345
+ ```
346
+
347
+ 1. **Auth** — a virtual key (`bir_…`); real credentials never leave the process.
348
+ 2. **Admit** — estimate cost; if no account could *ever* serve it, fail fast rather than
349
+ queue forever.
350
+ 3. **Route** — filter to eligible accounts, score, pick.
351
+ 4. **Lease** — atomically reserve the estimate so concurrent requests can't collectively
352
+ overshoot one account into a 429.
353
+ 5. **Proxy** — stream bytes through untouched.
354
+ 6. **Reconcile** — parse limit headers and usage; update the mirror, release the lease.
355
+ 7. **Recover** — classify the failure, transition the account, decide whether to retry.
356
+
357
+ ### Routing strategies
358
+
359
+ Set `routing.strategy`:
360
+
361
+ | Strategy | Behaviour |
362
+ |---|---|
363
+ | `sticky_headroom` | **Default.** Affinity, then most headroom. |
364
+ | `headroom` | `sticky_headroom` with affinity off, and nothing else changed. Identical to the default for traffic that has no cache owner, so it never routes better. |
365
+ | `cost_tiered` | Drain cheap accounts first, spill upward. |
366
+ | `priority` | Strict ordered failover: primary, then backup. |
367
+ | `least_loaded` | Baseline. |
368
+ | `round_robin` | Baseline. Ignores every signal, on purpose — it's there to benchmark against. |
369
+
370
+ Weights are config, not code. See `routing.weights` in `tokenbiryani.example.yaml`.
371
+ `tokenbiryani strategies` lists what this install has, including any installed plugins:
372
+ a strategy can ship in its own package under the `tokenbiryani.strategies` entry point,
373
+ supplying either weights or its own scoring.
374
+
375
+ ### Failure handling
376
+
377
+ | Upstream | Account | Request |
378
+ |---|---|---|
379
+ | `429` | cooldown until `retry-after` | retry elsewhere |
380
+ | `529` | short cooldown | retry |
381
+ | `500`/`502`/`503` | error tick, breaker may trip | retry with jitter |
382
+ | `401`/`403` auth | **disabled**, operator alerted | retry elsewhere |
383
+ | `403` model denied | model marked unsupported here | retry elsewhere |
384
+ | `400` / `413` | none — not the account's fault | **returned as-is, never retried** |
385
+
386
+ Retrying a `400` across the whole pool is the classic amplification bug in gateways like
387
+ this. The taxonomy exists to prevent exactly that.
388
+
389
+ ### The one failure that can't be hidden
390
+
391
+ Once the first SSE byte reaches the client, transparent failover is impossible — the
392
+ client is already rendering. So the gateway buffers the upstream stream until the first
393
+ `content_block_delta` (or `retry.first_token_grace_seconds`). A failure **before** that
394
+ point is retried silently on another account. A failure **after** it arrives as an SSE
395
+ `error` frame and the request ends.
396
+
397
+ ---
398
+
399
+ ## Docker
400
+
401
+ ```bash
402
+ docker compose up -d # gateway + Redis + a mock Anthropic
403
+ docker compose down # stop it
404
+ ```
405
+
406
+ Nothing to set first. The stack comes up with an empty pool, which is what the
407
+ onboarding wizard is for: open `http://localhost:8787/console` and add your first
408
+ credential there. Until you do, `/healthz` answers 503 and the container reports
409
+ itself unhealthy — an honest answer, since a pool with no accounts can serve
410
+ nothing.
411
+
412
+ `./src` is bind-mounted and watched, so editing a file on the host restarts the
413
+ gateway in about a second; the console's HTML and CSS need only a browser refresh;
414
+ and `./tests` is mounted too, so `docker compose exec gateway pytest -q` runs the
415
+ suite against the running stack.
416
+
417
+ The image runs as a non-root user, carries a healthcheck wired to `/healthz`, and
418
+ publishes to `127.0.0.1` rather than your LAN. `docker compose` builds the Dockerfile's
419
+ `dev` target; the default `runtime` target is what ships — wheel only, no source, no
420
+ test dependencies.
421
+
422
+ > The keys in `docker-compose.yml` are development values in a public repository.
423
+ > Copy `.env.example` to `.env` and replace them before this touches anything real —
424
+ > in particular `TOKENBIRYANI_SECRET_KEY`, which encrypts stored account credentials
425
+ > and must outlive the container. [docs/deployment.md](docs/deployment.md) explains both.
426
+
427
+ ## Operating it
428
+
429
+ Open **`http://localhost:8787/console`**. Everything the gateway can do, it can do from
430
+ there:
431
+
432
+ - **Accounts** — add, name, test, rename, rotate, enable, disable and delete
433
+ credentials without editing a file. Anthropic API keys, Bedrock, Vertex, and Claude
434
+ subscriptions via a browser login. Credentials are encrypted at rest; accounts
435
+ declared in `tokenbiryani.yaml` render locked, because the file is yours.
436
+ - **Usage** — token, cost and cache-hit-rate charts over 1h / 24h / 7d / 30d, grouped
437
+ by account, model or virtual key, with a totals table. This history is persisted, so
438
+ it survives a restart. See [docs/usage.md](docs/usage.md).
439
+ - **Overview** — capacity horizon, live request feed, per-account meters.
440
+ - **Efficiency** — the four questions the meters cannot answer: will the quota last
441
+ the week, is the prompt cache even switched on, is one conversation running away,
442
+ and is the pool reserving output it never uses. See
443
+ [docs/pacing.md](docs/pacing.md).
444
+ - **Requests** — the routing inspector: why each request went where it did.
445
+ - **Connect a client** — the exact export lines for this gateway's address.
446
+ - **Keys** — mint and revoke virtual keys.
447
+
448
+ ![The Accounts screen: the pool, each account's state and where it was declared, its
449
+ cost tier, priority and spend, with Test and Edit on every
450
+ row](docs/images/console-accounts.png)
451
+
452
+ Click an account and it opens: its limits, what its meters can and cannot tell you,
453
+ where its credential came from, an error breakdown by class, and its own recent
454
+ requests.
455
+
456
+ ![An account detail panel: a subscription account explaining that it reports
457
+ rolling-window utilisation rather than per-window budgets, where its token file lives
458
+ and when it expires, then state, requests, failures, p95, cache hit and spend, over its
459
+ recent requests](docs/images/console-account-detail.png)
460
+
461
+ Settings shows what the gateway is *currently running* — strategy and price table, each
462
+ marked with where its value came from, and a Reload that re-reads the file.
463
+
464
+ ![The Settings screen: routing strategy and price table, both marked "from the file",
465
+ with the bundled table's date beside it and live counts of accounts, queue depth and
466
+ priced models](docs/images/console-settings.png)
467
+
468
+ It is one server-rendered page plus a stylesheet, inside the package — no build step, no
469
+ Node toolchain added to a `pipx install`. The shell carries no data and needs no key; it
470
+ asks for an admin key on first load and keeps it in that browser only.
471
+
472
+ ### Does it actually work against the real API?
473
+
474
+ The one thing no mock can tell you is whether Anthropic spells its rate-limit headers
475
+ the way the router expects. If it doesn't, the mirror stays empty, every account reads
476
+ as full, and routing quietly degrades to round-robin — shredding the prompt cache while
477
+ looking healthy.
478
+
479
+ ```bash
480
+ tokenbiryani doctor --api-key sk-ant-...
481
+ ```
482
+
483
+ One real request, `max_tokens=1`. It prints the headers the upstream actually returned
484
+ next to the nine the limit mirror looks for, and what the mirror parsed out of them.
485
+ Non-zero exit if anything is missing. Run it once after you first point this at
486
+ production.
487
+
488
+ Or stay in the terminal:
489
+
490
+ ```bash
491
+ tokenbiryani status # the pool, in the terminal you're already in
492
+ tokenbiryani status --json # same data, for scripts
493
+ ```
494
+
495
+ ```
496
+ POOL 702k tok ready · next reset 00:12 · queue 0 · cache 94%
497
+
498
+ ACCOUNT STATE REQ INPUT OUTPUT RESET CACHE
499
+ acct-01 ● ready 98% 82% 79% 00:41 97%
500
+ acct-02 ● cooling 4% 0% 6% 00:27 —
501
+
502
+ 1h 412 requests · 3 failovers · 1 cache break · 0 errors · $18.40
503
+ ```
504
+
505
+ | Endpoint | |
506
+ |---|---|
507
+ | `GET /console` | the operator console |
508
+ | `POST /v1/messages` | Messages API, streaming and not |
509
+ | `POST /v1/messages/count_tokens`, `GET /v1/models` | passthrough |
510
+ | `GET /healthz` | 200 while any account is ready |
511
+ | `GET /admin/status` | pool snapshot |
512
+ | `GET /admin/usage` | bucketed usage history for the charts |
513
+ | `POST /admin/accounts` · `PATCH` · `DELETE` · `POST /admin/accounts/{id}/test` | manage credentials at runtime |
514
+ | `POST /admin/keys` · `DELETE /admin/keys/{name}` | mint and revoke keys at runtime |
515
+ | `GET /admin/accounts/{id}` | one account: limits, error breakdown by class, its own recent requests |
516
+ | `POST /admin/reload` | re-read the config file |
517
+ | `GET /admin/requests/{id}` | **why that request went where it did** — attempt chain, per-candidate scores, verdicts |
518
+ | `GET /admin/horizon` | projected capacity for the next hour |
519
+ | `GET /admin/estimation` | what the output estimator has learned, per model |
520
+ | `GET /admin/pacing` | **is this pool on course to spend its week** — or run dry early, or strand quota |
521
+ | `GET /admin/sessions` | the most expensive conversations in the window, runaways flagged |
522
+ | `GET /admin/cache-advice` | **why the cache hit rate is what it is** — whether the client ever sent a breakpoint, per key and model |
523
+ | `GET /admin/events` | live SSE feed |
524
+
525
+ The request inspector is the point. `filtered — cooling, 27s remaining` is a complete
526
+ answer; "load balanced" is not.
527
+
528
+ **`/admin/*` requires a key with `admin: true`.** It exposes account ids, spend and key
529
+ management, so a tenant key must not reach it. A gateway with no keys configured at all
530
+ is loopback development mode and stays fully open.
531
+
532
+ ### Managing keys at runtime
533
+
534
+ ```bash
535
+ curl -sX POST localhost:8787/admin/keys -H "x-api-key: $ADMIN_KEY" \
536
+ -d '{"name":"tenant-1","pool":["acct-02"],"rpm":60,"spend_cap_usd":5}'
537
+ # -> {"key": "bir_...", "record": {...}} the plaintext appears exactly once
538
+
539
+ curl -sX DELETE localhost:8787/admin/keys/tenant-1 -H "x-api-key: $ADMIN_KEY"
540
+ ```
541
+
542
+ ![The Keys screen: a form minting a key with pool, rpm, spend cap, priority and admin
543
+ flag, above a table of existing keys marked "in config" with their pools and
544
+ limits](docs/images/console-keys.png)
545
+
546
+ Minted keys are stored **hashed**, so a leaked state store is not a leaked key, and they
547
+ live in the shared store — one instance honours a key another minted. Keys declared in
548
+ the config file belong to the file: the API will not revoke them.
549
+
550
+ ### The spill lane
551
+
552
+ With `batch.enabled: true`, a `batch`-priority request that finds the pool saturated goes
553
+ to the **Message Batches API** instead of waiting. The gateway holds the connection while
554
+ it polls, bounded by that request's own wait budget. A batch that outlives the budget is
555
+ **cancelled upstream** and its id returned in `x-tokenbiryani-batch-id`, so nothing is
556
+ silently abandoned. If submission fails the request falls back to the normal queue — the
557
+ spill lane is an optimisation, never a dependency. Streaming requests never spill.
558
+
559
+ ### Account types
560
+
561
+ | `type` | Notes |
562
+ |---|---|
563
+ | `anthropic_api` | Anthropic API keys. The default. |
564
+ | `bedrock` | AWS Bedrock. SigV4-signed; its binary event-stream is decoded back to SSE so the rest of the gateway sees ordinary streaming. Needs `pip install "tokenbiryani[bedrock]"`. |
565
+ | `vertex` | Google Vertex AI. Bearer token from application-default credentials; returns real SSE already. Needs `pip install "tokenbiryani[vertex]"`. |
566
+ | `oauth` | A Claude subscription (Max/Pro). Three token sources, offered in the console's Add-account dialog: this machine's Claude Code login (nothing to configure), a long-lived token from `claude setup-token`, or an OAuth login the gateway runs itself — that last one stays disabled until you supply the provider endpoints, which this project will not guess at. Read [docs/oauth.md](docs/oauth.md) first: a subscription reports rolling-window utilisation rather than per-window budgets, so it keeps headroom routing and failover but has no leases, no admission control and no capacity horizon. |
567
+
568
+ ![The Add-account dialog for a Claude subscription: three token sources — this
569
+ machine's Claude Code login, a long-lived token, or an OAuth login — with the machine
570
+ scan listing each credentials file it found, its plan, and whether it is current or
571
+ stale](docs/images/console-subscription-session.png)
572
+
573
+ All four sit in one pool, so a request can fail over from an API key to Bedrock. Use
574
+ `options.model_map` to translate your callers' model names into each platform's ids.
575
+
576
+ ```yaml
577
+ accounts:
578
+ - id: acct-01
579
+ type: anthropic_api
580
+ api_key: ${ANTHROPIC_API_KEY}
581
+ - id: acct-bedrock
582
+ type: bedrock
583
+ cost_tier: 1.2
584
+ options:
585
+ region: us-east-1
586
+ model_map:
587
+ claude-test-1: anthropic.claude-3-5-sonnet-20241022-v2:0
588
+ - id: acct-vertex
589
+ type: vertex
590
+ options:
591
+ project: my-project
592
+ region: us-central1
593
+ ```
594
+
595
+ ### Request headers
596
+
597
+ | Header | |
598
+ |---|---|
599
+ | `X-TokenBiryani-Session` | pin a conversation to one affinity key instead of the fingerprint |
600
+ | `X-TokenBiryani-Priority` | `interactive` (default) or `batch`. Batch traffic yields the queue to interactive traffic when the pool is saturated |
601
+ | `X-TokenBiryani-Max-Wait` | seconds this request will wait for capacity. A client can shorten its own budget but never extend it past the operator's ceiling |
602
+
603
+ Per-key defaults for the last two live under `keys:` as `priority` and `max_wait_seconds`.
604
+
605
+ ### Spend caps
606
+
607
+ Caps come at two scopes. `keys[].spend_cap_usd` bounds everything a key does;
608
+ `session_cap_usd` and `session_max_turns` bound one *conversation* inside that
609
+ allowance. The second exists because the first cannot see a runaway: one agent loop
610
+ resending a large prefix a few hundred times is a whole key's cap with that key's
611
+ name on it, and until it trips there is nothing to look at. `GET /admin/sessions`
612
+ ranks conversations by cost and flags the ones past `sessions.runaway_turns`.
613
+
614
+ Caps are **windowed, not lifetime** (`spend.window_hours`, default 24). A lifetime cap on
615
+ a persistent store would eventually wedge the gateway shut and stay that way.
616
+
617
+ They only survive a restart if the store does. `store.backend: memory` (the default)
618
+ loses affinity and the spend ledger when the process dies — meaning every cap silently
619
+ resets. Use `sqlite` for a real deployment:
620
+
621
+ ```yaml
622
+ store:
623
+ backend: sqlite
624
+ path: tokenbiryani.db
625
+ ```
626
+
627
+ ### Running more than one instance
628
+
629
+ `store.backend: redis` shares affinity, the spend ledger and per-key rate counters
630
+ between processes. Without it two instances each keep their own affinity map — so a
631
+ conversation ping-pongs between them and loses its cache — and each enforces its own
632
+ half of every spend cap.
633
+
634
+ ```yaml
635
+ store:
636
+ backend: redis
637
+ url: redis://127.0.0.1:6379/0
638
+ namespace: tokenbiryani
639
+ ```
640
+
641
+ Needs the optional dependency: `pip install "tokenbiryani[redis]"`.
642
+
643
+ ### Costs
644
+
645
+ Costs are reported, and spend caps enforced, only for models that have a price.
646
+ `pricing: builtin` takes the dated table that ships with this release — the console
647
+ shows its date beside every cost — and naming a model under `pricing:` overrides it.
648
+
649
+ Prices live in a dated data file rather than in code, because the gateway must never
650
+ bill you against a number nobody can attribute. A dated file whose date is on screen
651
+ can be attributed; a dict compiled into a release cannot.
652
+
653
+ ### Privacy and security
654
+
655
+ - Prompt bodies are **never logged**. Only accounting: ids, accounts, tokens, latency.
656
+ - Credentials are never logged and are masked in every admin response.
657
+ - Binds to loopback. Refuses to start on a public interface without both
658
+ `server.allow_remote: true` and at least one configured key.
659
+
660
+ ---
661
+
662
+ ## Documentation
663
+
664
+ Full docs build from `docs/` with `mkdocs serve`. Start with
665
+ [Why is my bill higher?](docs/caching.md) — it is the page that changes how you
666
+ configure this thing.
667
+
668
+ ## Status
669
+
670
+ All eight milestones in `PLAN.md` are implemented and tested; see `TODO.md` for the
671
+ table and for what was deliberately left out.
672
+
673
+ Working today: passthrough and streaming, multi-account pooling, the error taxonomy,
674
+ retry and failover, the rate-limit mirror, token estimation and leases, headroom scoring,
675
+ circuit breakers, session affinity and cache accounting, admission control and a bounded
676
+ priority queue, virtual keys with model/pool/rpm/spend scoping,
677
+ structured logs, config hot reload, the admin API, and the CLI. Request priority with a
678
+ per-request wait budget, a batch spill lane, and SQLite-backed persistence for affinity
679
+ and windowed spend, a Redis store for multi-instance deployments, and runtime key
680
+ management behind an admin boundary, plus Bedrock and Vertex adapters and pluggable
681
+ routing strategies, and the operator console. 379 tests including a lease-concurrency
682
+ suite, a reproducible benchmark, and an end-to-end smoke test over real sockets
683
+ (`scripts/smoke.sh`).
684
+
685
+ All eight milestones in `PLAN.md` are built. `docs/UI-DESIGN.md` is the console's design
686
+ brief, and the console follows it.
687
+
688
+ **Credential types.** Anthropic API keys are the supported path. A *subscription*
689
+ account works — routing your own subscription through your own local gateway is the
690
+ ordinary case, and it reports enough (rolling-window utilisation) to route on — but
691
+ pooling several so their limits add up runs against Anthropic's consumer terms. Leases,
692
+ admission control and the capacity horizon need absolute token counts and stay dark for
693
+ those accounts. See [docs/oauth.md](docs/oauth.md).
694
+
695
+ ## License
696
+
697
+ Apache-2.0.