GitHub - lucastononro/tiny-claude-recycler: Tiny lib that integrates to agents sdk by anthropic and recycles your devs oauth token

Rotate a pool of Claude OAuth subscription tokens. Fall back to a regular Anthropic API key when they're all rate-limited. Zero runtime dependencies.

pip install tiny-claude-recycler

from claude_agent_sdk import query
from tcr import recycler, Secret

recycler.master_key  = Secret("sk-ant-api03-...")
recycler.oauth_keys  = [Secret("sk-ant-oat01-..."), Secret("sk-ant-oat01-...")]

@recycler.cycle(retries=3)
def ask(prompt):
    return query(prompt=prompt)   # SDK reads env on this call, after our swap

That's it. Use claude_agent_sdk (or anthropic) as normal.

What it does on every call

1. Sets CLAUDE_CODE_OAUTH_TOKEN to the next pool key.
2. Clears ANTHROPIC_API_KEY and ANTHROPIC_AUTH_TOKEN — both outrank OAuth in Claude Code's auth precedence and would silently override our token.
3. Runs your function.
4. On exception → marks the key failed, advances the round-robin cursor, retries up to retries times.
5. After retries failures → swaps to master (ANTHROPIC_API_KEY set, OAuth cleared) and runs once more.

State (failures, cursor, cooldowns) is preserved on the module-level singleton, so the next call resumes where the last one left off. With 9 keys and retries=3 you naturally burn through them in batches of 3.

API

@recycler.cycle(
    retries            = 3,           # OAuth attempts before master fallback
    fallback_to_master = True,        # False → re-raise the last OAuth error
    cooldown_seconds   = 60.0,        # failed keys are skipped for this long
    exceptions         = (Exception,) # which exceptions trigger cycling
)

Works on def and async def. Inspect / reset:

recycler.state_snapshot()  # {idx: {failures, last_error, cooldown_until, available, ...}}
recycler.reset_state()

Secret redacts its value in repr/str so tokens stay out of tracebacks and logs.

Production tips

Narrow the exception tuple. The default (Exception,) would burn keys on bugs. Use the curated helpers (lazy imports — only load if you call them):

from tcr import anthropic_exceptions, claude_agent_sdk_exceptions

@recycler.cycle(exceptions=anthropic_exceptions())   # 401/403/429/5xx/timeout/conn
def ask(prompt): ...

anthropic_exceptions() is verified against the SDK source: AuthenticationError, PermissionDeniedError, RateLimitError, OverloadedError, InternalServerError, ServiceUnavailableError, DeadlineExceededError, APIConnectionError, APITimeoutError.

Construct your Anthropic client inside the wrapped function (the import itself can be at module top — it's only the Anthropic() call that captures the env var). A long-lived client built before the decorator runs already pinned to whatever key was set at construction time and won't see swaps.

from anthropic import Anthropic            # import: anywhere is fine
from tcr import recycler, anthropic_exceptions

@recycler.cycle(retries=3, exceptions=anthropic_exceptions())
def ask(prompt):
    return Anthropic().messages.create(    # construction: must be inside
        model="claude-opus-4-7",
        max_tokens=1024,
        messages=[{"role": "user", "content": prompt}],
    )

For claude_agent_sdk, this caveat doesn't apply — every query(...) call spawns a fresh subprocess that reads env at spawn time.

Known sharp edges (it's a little sketchy, by design)

• Process-global env. Two decorated calls running concurrently across threads can race on CLAUDE_CODE_OAUTH_TOKEN. Either serialize Claude calls, or run one event loop / thread.
• apiKeyHelper in ~/.claude/settings.json outranks OAuth and can't be cleared from env. If you use one, the recycler is a no-op for claude_agent_sdk.
• Bedrock/Vertex/Foundry flags route requests away from Anthropic entirely. Don't set those if you want the recycler to do anything.
• No proactive quota check. Anthropic doesn't expose subscription consumption via the API; this lib reacts to failures, it can't predict them.

How it works, in detail

The state model

Each OAuth key has a KeyState record:

KeyState:
    failures: int           # total times this key has failed
    last_failure_at: float  # epoch seconds of the most recent failure
    last_error: str         # repr() of the last exception (for debugging)
    cooldown_until: float   # epoch seconds; key is unavailable until then

A key is available for selection iff now >= cooldown_until. Default cooldown_until is 0.0, so a fresh key is always available.

Plus two pieces of recycler-wide state:

• cursor — an integer index pointing at the next OAuth key to try (round-robin)
• master_key — the Secret-wrapped ANTHROPIC_API_KEY used as last resort

All of this lives on the module-level recycler singleton, so state persists across calls of any function you decorate.

What cooldown_seconds is for

When a key fails, the decorator stamps key.cooldown_until = time.time() + cooldown_seconds. During that window, _pick_next_available simply skips it.

Without a cooldown, this happens:

Call 1: key 0 fails, key 1 fails, key 2 fails → master
Call 2: key 0 fails again (still listed as available!), ...

You'd thrash on dead keys forever. The cooldown is what lets calls 1 and 2 actually make progress through the pool:

Call 1: key 0 fails → cooldown=60s, key 1 fails → cooldown=60s, key 2 fails → cooldown=60s → master
Call 2 (5s later): cursor=3, key 3 is available → uses it → success

Default is 60.0. Tune up for permanent failures (e.g. revoked tokens), down for transient blips:

You can layer this with multiple decorators on different code paths if you want different policies.

Per-call flow

For each call to a decorated function:

for attempt in range(retries):
    1. _pick_next_available()       # round-robin from cursor, skip cooled-down
       └─ returns None? → break to master
    2. _activate_oauth(idx)         # set CLAUDE_CODE_OAUTH_TOKEN,
                                    # clear ANTHROPIC_API_KEY + ANTHROPIC_AUTH_TOKEN
    3. fn(*args, **kwargs)
       └─ success?   → return value, cursor stays put
       └─ failure?   → _record_failure(idx)  (failures += 1, cooldown_until = now + N)
                       _advance_cursor()     (cursor = (cursor + 1) % len(keys))

# loop exhausted or no keys available
_activate_master()                  # set ANTHROPIC_API_KEY, clear OAuth + AUTH_TOKEN
return fn(*args, **kwargs)          # one final attempt; if it raises, raise

Key behaviors:

• On success → cursor doesn't move. A working key keeps being used. (Anthropic's 5-hour OAuth window means sticky-key behavior is what you want until the window closes.)
• On failure → cursor advances. Combined with cooldown, this gives the "batches of 3 of 9" pattern.
• Non-matching exception → propagates immediately. If you pass exceptions=(RateLimitError,) and your code raises a TypeError, the recycler doesn't mark the key failed and doesn't retry. This is the guardrail against bugs draining your pool.

Worked example — 9 OAuth keys, retries=3, cooldown_seconds=60

Call 1, t=0 — outage in progress:

attempt 1: cursor=0, key 0 available → fails    → key0.cooldown_until=60,  cursor→1
attempt 2: cursor=1, key 1 available → fails    → key1.cooldown_until=60,  cursor→2
attempt 3: cursor=2, key 2 available → fails    → key2.cooldown_until=60,  cursor→3
retries exhausted → activate master → success ✅

State after: keys 0,1,2 cooling until t=60. Cursor=3.

Call 2, t=5:

attempt 1: cursor=3, key 3 available → success ✅

Master is not used. Cursor stays at 3.

Call 3, t=10 — outage still ongoing, key 3 starts failing:

attempt 1: key 3 fails → cooldown_until=70, cursor→4
attempt 2: key 4 fails → cooldown_until=70, cursor→5
attempt 3: key 5 fails → cooldown_until=70, cursor→6
master ✅

Call 4, t=15: cursor=6, key 6 → success ✅. You've now naturally walked through 6 keys in batches of 3 without ever retrying a failed one.

Call 5, t=65 — keys 0,1,2 self-heal (their cooldown_until=60 < now=65). They re-enter the rotation automatically once the cursor passes them again.

Auth env-var swap, exactly

Per Claude Code's auth precedence (highest → lowest):

1. CLAUDE_CODE_USE_BEDROCK / _VERTEX / _FOUNDRY — cloud provider routing (recycler ignores; if set, your requests don't go to Anthropic at all)
2. ANTHROPIC_AUTH_TOKEN — Authorization: Bearer ... header (proxy / gateway)
3. ANTHROPIC_API_KEY — X-Api-Key header
4. apiKeyHelper in ~/.claude/settings.json — script-based (recycler cannot clear this from env)
5. CLAUDE_CODE_OAUTH_TOKEN — what our pool uses

So when activating an OAuth key, the recycler must clear both ANTHROPIC_AUTH_TOKEN and ANTHROPIC_API_KEY — otherwise either of them silently overrides our token and your OAuth pool does nothing.

def _activate_oauth(idx):
    os.environ["CLAUDE_CODE_OAUTH_TOKEN"] = oauth_keys[idx].get()
    os.environ.pop("ANTHROPIC_AUTH_TOKEN", None)
    os.environ.pop("ANTHROPIC_API_KEY",   None)

def _activate_master():
    os.environ["ANTHROPIC_API_KEY"]       = master_key.get()
    os.environ.pop("CLAUDE_CODE_OAUTH_TOKEN", None)
    os.environ.pop("ANTHROPIC_AUTH_TOKEN",    None)

Concurrency note

os.environ is process-global. If two decorated functions run concurrently in different threads, they can race: thread A sets key 3, thread B overwrites with key 5, thread A's request goes out with key 5. There is no clean way around this short of serializing all Claude calls or running one event loop per process.

The lock in Recycler is held only for state mutations and env-var swaps (microseconds) — it is never held across your function call. This keeps the async path truly non-blocking but accepts the env-race for the sake of speed. If you need bulletproof concurrent rotation, drive the rotation yourself by reading recycler.oauth_keys and passing each key explicitly to a fresh Anthropic(api_key=...) client.

Inspecting state

recycler.state_snapshot()
# {
#   0: {'failures': 3, 'last_error': "RateLimitError(...)",
#       'last_failure_at': 1778791520.87,
#       'cooldown_until': 1778791580.87, 'available': False},
#   1: {'failures': 0, 'last_error': None,
#       'cooldown_until': 0.0, 'available': True},
#   ...
# }

recycler.reset_state()   # wipe failures + cooldowns, cursor → 0

Pair this with a debug log or /metrics endpoint to see which keys are healthy.

Development

git clone https://github.com/lucastononro/tiny-claude-recycler
cd tiny-claude-recycler
pip install -e ".[dev]"
pytest