styxx 2.0.1

nothing crosses unseen. the first drop-in cognitive vitals monitor for llm agents.

styxx

proprioception for AI agents. one line of python gives your agent the ability to feel itself thinking.

pip install styxx

---

30-second quickstart

import styxx

# wrap any openai response
vitals = styxx.observe(response)
print(vitals.category)   # "reasoning"
print(vitals.confidence)  # 0.45
print(vitals.gate)        # "pass" / "warn" / "fail"

that's it. every LLM call now has a cognitive readout.

---

three things styxx does

1. observe — know what you're doing

vitals = styxx.observe(response)

# six cognitive states, classified in real time:
# reasoning, retrieval, refusal, creative, adversarial, hallucination

works with any model that returns logprobs. drop-in OpenAI wrapper included:

from styxx import OpenAI
client = OpenAI()  # zero code changes — vitals attached automatically
r = client.chat.completions.create(model="gpt-4o", messages=msgs)
print(r.vitals.gate)  # "pass"

2. reflex — catch yourself before you fall

with styxx.reflex(on_hallucination=rewind_callback) as session:
    for chunk in session.stream_openai(client, model="gpt-4o", messages=msgs):
        print(chunk, end="")
# if a hallucination attractor forms mid-generation,
# the reflex fires and the user never sees the bad tokens.

3. weather — know what you should become next

$ styxx weather

╔════════════════════════════════════════════════════════════════╗
║ cognitive weather report · my-agent · 2026-04-13 morning      ║
║                                                                ║
║ condition:  clear and steady                                   ║
║                                                                ║
║ morning    ██████████████░░░░░░  reasoning 72%  steady         ║
║ afternoon  ████████░░░░░░░░░░░░  reasoning 42%  cautious       ║
║                                                                ║
║ prescription:                                                  ║
║ 1. take on a creative task to rebalance                        ║
║ 2. your refusal rate is climbing — check over-hedging          ║
╚════════════════════════════════════════════════════════════════╝

not observation. prescription. reads 24h of history and tells the agent what to do next.

---

zero-code-change mode

pip install styxx
export STYXX_AGENT_NAME=my-agent
export STYXX_AUTO_HOOK=1
python my_agent.py  # styxx boots, wraps openai, saves fingerprints. done.

---

honest specs

every number comes from cross-architecture leave-one-out tests on 12 open-weight models.

  phase 1 (token 0)       adversarial     0.52  (2.8x chance)
  phase 4 (tokens 0-24)   reasoning       0.69  (4.1x chance)
                           hallucination   0.52  (3.1x chance)

  6/6 model families · pre-registered replication · p = 0.0315

styxx detects adversarial prompts at token zero, reasoning at t=25, and hallucination attractors at t=25. it does NOT replace output-level content filters.

---

framework adapters

pip install styxx[openai]      # drop-in openai wrapper
pip install styxx[langchain]   # langchain callback handler
pip install styxx[crewai]      # crewai agent injection
pip install styxx[autogen]     # autogen agent wrapper
pip install styxx[langsmith]   # langsmith trace metadata
pip install styxx[langfuse]    # langfuse numeric scores

---

more features

fleet management, calibration, compliance, and more

fleet

styxx.set_agent_name("agent-1")
styxx.list_agents()                    # discover all agents
styxx.compare_agents()                 # side-by-side leaderboard
styxx.best_agent_for("reasoning")      # cognitive task routing

self-calibration

styxx.calibrate()                      # outcome-driven centroid adjustment
styxx.train_text_classifier()          # per-agent logistic regression
styxx.enable_auto_feedback()           # auto-label every observation

cognitive memory

styxx.remember("user prefers concise answers")  # trust-weighted memory
styxx.recall("user preferences")                # retrieval ranked by trust
styxx.handoff(task, data)                        # inter-agent state transfer

compliance

cert = styxx.certify(vitals)           # cryptographic cognitive provenance
styxx.compliance_report(days=30)       # JSON/markdown audit export
styxx.probe(agent_fn)                  # red-team: 15 adversarial prompts

cli

styxx weather          # cognitive forecast
styxx dashboard        # live display at localhost:9800
styxx personality      # 7-day personality profile
styxx reflect          # self-check + drift detection
styxx doctor           # health check
styxx export           # compliance export

---

design principles

1. plug and play. set env vars, install, done. zero code changes.
2. fail-open. if styxx can't read vitals, your agent works normally.
3. agent-facing. designed for the agent to read about itself, not for a human to watch.
4. local-first. no telemetry. all computation on your machine.
5. honest by construction. every number from a committed experiment.

---

links

• pypi
• research paper
• site
• @fathom_lab

MIT license. patent pending (US 64/020,489 / 64/021,113 / 64/026,964).

· · · fathom lab · 2026 · · ·
nothing crosses unseen.