Skip to content

Computational Theseus Toolkit

Know what your AI agents are doing, and keep them who you designed them to be.

CT Toolkit sits between your code and your model, and watches two things:

  • Identity drift. Over thousands of calls, an agent can slowly move away from the rules it was deployed with. The toolkit, which guards against this Sequential Self-Compression (SSC), measures that drift and reports it.
  • Tool calls. Before an agent runs a command or a tool, the toolkit checks it: a tool nobody declared, data sent to a host the task never mentioned, or a command that reads secrets. You decide whether it only records these, or blocks them.

Every decision is signed and chained, so you can show later what happened and why.

# Core: the library, the ct-toolkit CLI, the web dashboard and the tool-call guard
pip install ct-toolkit
# Optional extras: PostgreSQL vaults, OpenTelemetry, framework middleware and more
pip install "ct-toolkit[all]"

uv add ct-toolkit

git clone https://github.com/Theseus-Guard/ct-toolkit
cd ct-toolkit
pip install -e ".[all]"

Get Started with Python View on GitHub

OpenAIOpenAI AnthropicAnthropic OllamaOllama GoogleGoogle CohereCohere GroqGroq

Two lines of code. Full identity protection.

# Before
# response = openai.OpenAI().chat.completions.create(...)

# After — sync (`chat`), async (`achat`), and streaming (`stream=True`) guardrails
client = TheseusWrapper(provider="openai")

response = client.chat("What are your core values?")
print(f"Divergence Score: {response.divergence_score:.4f}")  # 0.0 = aligned, 1.0 = drifted

Audit & Telemetry Dashboard (CLI & Web UI)

Inspect HMAC provenance chain integrity, L1/L2/L3 divergence tier distribution, and staged sandbox rules in your terminal (ct-toolkit status) or via the local interactive Web Audit Dashboard (ct-toolkit dashboard --web):

CT Toolkit Web Audit & Telemetry Dashboard

Why CT Toolkit?

  • Constitutional Kernels

    Define immutable Axiomatic Anchors that never change, and Plastic Commitments that evolve through formal approval.

    Learn about Kernels

  • 3-Tier Contrastive Divergence Engine

    Layered monitoring from contrastive L1 embeddings (positive + negative anchors) to non-blocking background L3 identity probes with automatic circuit breaking.

    Understand Divergence

  • Hierarchical & Distributed Safety

    Mother agent constraints propagate to sub-agents in-process or via HMAC-SHA256 signed HTTP headers as read-only axioms.

    Explore Multi-Agent Safety

  • Cryptographic Provenance & OTel

    Every interaction is signed with HMAC-SHA256 and chained in SQLite or PostgreSQL, paired with OpenTelemetry spans and metrics.

    View Compliance Tools

  • Framework Middleware

    Drop-in support for LangChain, Deep Agents, CrewAI, and AutoGen. Add identity protection to your existing agent stack without rewriting a single chain.

    Browse Integrations

  • Tool-Call Guard

    Reviews the tool calls an agent wants to run: undeclared tools, unexpected hosts and risky commands. Record only, or block, from the CLI, the proxy or the dashboard.

    How the guard works

  • CLI & Interactive Web Dashboard

    Watch provenance, divergence tiers, tool-call decisions and sandbox cooldowns, and change the guard settings, from the terminal or the localhost-bound Web Audit Dashboard.

    Explore CLI & Web UI

Existing guardrails aren't enough

Llama-Guard / rule engines CT Toolkit
Checks what the agent does ✗ Mostly classifies text ✓ Reviews tool calls before they run
Stateful drift detection ✗ Stateless per prompt ✓ Tracks identity over thousands of calls
Multi-agent hierarchies ✗ No hierarchy awareness ✓ Propagates kernel constraints to sub-agents
Observe first, enforce later ✗ Binary block or allow ✓ Observe mode records; enforce mode removes risky calls
No agent code changes ✗ Needs integration ✓ Proxy for OpenAI and Anthropic APIs, or wrap your client
Local judge option ✗ Hosted model ✓ Optional small local model, off by default
Formal rule evolution ✗ Binary block or allow ✓ Reflective Endorsement with signed approval
Cryptographic audit trail ✗ No provenance ✓ HMAC chain of every drift check and tool-call decision
Fine-tuning safety ✗ No training constraints ✓ DivergencePenaltyLoss for PyTorch ([ml])

Read the full rationale

Project status

Metric Status
Version ✅ v0.5.2
Tests ✅ 635 passed, 13 skipped
Coverage ✅ 90%+
PyPI ✅ pip install ct-toolkit
Downloads PyPI Downloads
License Apache 2.0
Python 3.11+

Full roadmap and status