leancontext 2.1.0

Trim the tool output your AI agent re-sends every turn. Keep the signal, drop the noise.

AI agents re-send every tool result (logs, JSON, diffs, stack traces, HTML) to the model on every turn, and most of it is redundancy you pay for again and again. LeanContext sits between your agent and the model and reduces those payloads to their signal: deterministically, with a fidelity score on every reduction, and without ever breaking the agent.

from leancontext import reduce

@reduce
def search_logs(query: str) -> str:
    return run_log_search(query)   # ~10k tokens of logs in, ~1k out, error lines kept

See it

$ python bench.py
sample              kind          before   after  saved  fidelity
-----------------------------------------------------------------
log (incident)      log            52642     100   100%      100%
json (RAG chunks)   json            1862    1391    25%      100%
html (web fetch)    html            1672    1093    35%      100%
diff (patch)        diff             639      81    87%      100%
stacktrace          stacktrace       896      94    90%      100%
-----------------------------------------------------------------
TOTAL                              57711    2759    95%

Counts above use the built-in heuristic tokenizer (≈4 chars/token). Install the tiktoken extra for exact model token counts — the ratios are similar (~92% on this sample). The reduced text is identical either way.

A real incident log, before and after:

# before  (902 lines)
2026-06-21T09:00:01Z INFO  [gateway] req id=a1 path="/v1/render" status=200 ms=12
... 900 near-identical INFO lines ...
2026-06-21T09:10:43Z FATAL [render] OOM killed worker=7 doc="deck-8842" root cause

# after
2026-06-21T09:00:01Z INFO  [gateway] req id=a1 path="/v1/render" status=200 ms=12   ⟪×900 similar⟫
2026-06-21T09:10:43Z FATAL [render] OOM killed worker=7 doc="deck-8842" root cause

The redundant lines collapse to a count. The FATAL line that explains the crash is kept intact.

Why it works

The model API is the bulk of an agent's cost, and most of that is input tokens. A tool result added on one turn is re-sent on every later turn, so the bill grows with the length of the conversation, not just the work done. Those payloads are mostly repetition. LeanContext keeps the errors, anomalies, and identifiers, and collapses the rest.

How it compares

Install

pip install leancontext                  # core, standard library only
pip install "leancontext[integrations]"  # openai, anthropic, litellm, fastapi adapters
pip install "leancontext[otel]"          # OpenTelemetry metrics
pip install "leancontext[mcp]"           # MCP server
pip install "leancontext[tiktoken]"      # exact token counts (used automatically when present)

Use it

Three levels, one core. Every path fails open: if anything goes wrong, you get the original text back.

import leancontext

clean = leancontext.reduce(tool_output).text     # 1) manual

@leancontext.reduce                              # 2) decorator, one line per tool
def search_logs(q: str) -> str:
    ...

tools  = leancontext.wrap(tools)                 # 3) wrap all tools, or an SDK client
client = leancontext.wrap(openai_client)         #    (wrap_anthropic / wrap_gemini too)

Every reduction is inspectable:

r = leancontext.reduce(tool_output)
r.text                            # what to send to the model
r.tokens_before, r.tokens_after
r.ratio                           # fraction saved
r.fidelity                        # 0..1 signal preserved

Integrations

CI exercises OpenAI (chat + Responses), Anthropic, LiteLLM, the standalone proxy, OpenTelemetry, and the MCP server against the real packages. Message reduction for all formats (including Gemini) is unit-tested directly. The framework adapters (LangChain / LangGraph / Agno) and the SDK-level Gemini client wrapper are provided best-effort and are not yet covered in CI against the live SDKs.

Reducers

Anything else, or any payload below the size, saving, or fidelity thresholds, passes through unchanged.

How it works

Each tool output flows through fail-open gates (hash, size check, type detection, the typed reducer, then a saving and fidelity check) and returns either the reduced text or the original. Results are cached by content hash, so a payload re-sent across turns is reduced only once. See docs/ARCHITECTURE.md for diagrams.

Cost and telemetry

from leancontext.cost import CostTracker

tracker = CostTracker(model="claude-sonnet-4-6").install()
# ... run your agent ...
tracker.report()    # {tokens_saved, usd_saved, ratio, cache_safe: True}

Configuration

leancontext.disable()                         # global kill switch (or env LEANCONTEXT_DISABLED=1)
leancontext.reduce(x, min_saving=0.1, min_fidelity=0.85)
leancontext.on_reduction(callback)            # telemetry hook (composable)
leancontext.use_tiktoken("gpt-4o")            # force a specific model's tokenizer

Roadmap

CI-verified LangChain / LlamaIndex / CrewAI / Agno adapters, accurate provider tokenizers by default, and broader Anthropic native interop.

Contributing

Issues and PRs welcome. Run pytest. Reducers are pure functions, str -> (reduced, notes), and must be deterministic and value-preserving. See AGENTS.md for the design rules.

License

Apache-2.0