langchain-kredo 0.1.0

LangChain integration for the Kredo agent attestation protocol

langchain-kredo

LangChain integration for the Kredo agent attestation protocol.

Check agent trust, search attestations, enforce policy, and create signed attestations — all from your LangChain pipeline.

Install

pip install langchain-kredo

Quick Start

from langchain_kredo import KredoSigningClient, KredoTrustGate

# Initialize with Ed25519 signing key (hex seed)
client = KredoSigningClient(
    signing_key="your-hex-seed",
    name="my-pipeline",
    agent_type="agent",
)

# Check trust before delegating to an agent
gate = KredoTrustGate(client, min_score=0.3, block_warned=True)
result = gate.check("ed25519:agent-pubkey-here")
if result.passed:
    print(f"Agent trusted (score: {result.score})")

Components

KredoSigningClient

Wraps the Kredo Discovery API with Ed25519 signing. Read operations query api.aikredo.com. Write operations build Pydantic models, sign with your key, and submit.

client = KredoSigningClient(
    signing_key=sk,           # SigningKey, bytes, hex string, or env var
    name="my-agent",
    agent_type="agent",       # "agent" or "human"
)

# Read
profile = client.get_profile("ed25519:...")
results = client.search(domain="security", min_proficiency=3)
taxonomy = client.get_taxonomy()

# Write (requires signing key)
client.register()
client.attest_skill(
    subject_pubkey="ed25519:...",
    domain="security",
    skill="vulnerability_assessment",
    proficiency=4,
    context="Demonstrated expert-level vuln assessment in CTF",
)

Key resolution order: signing_key param > KREDO_PRIVATE_KEY env var > read-only mode.

LangChain Tools

Four tools for agent pipelines:

from langchain_kredo import (
    KredoCheckTrustTool,
    KredoSearchAttestationsTool,
    KredoSubmitAttestationTool,
    KredoGetTaxonomyTool,
)

tools = [
    KredoCheckTrustTool(client=client),
    KredoSearchAttestationsTool(client=client),
    KredoSubmitAttestationTool(client=client),
    KredoGetTaxonomyTool(client=client),
]

Tool	Name	Purpose
KredoCheckTrustTool	kredo_check_trust	Check agent reputation + skills + warnings
KredoSearchAttestationsTool	kredo_search_attestations	Find agents by skill/domain/proficiency
KredoSubmitAttestationTool	kredo_submit_attestation	Sign and submit skill attestation
KredoGetTaxonomyTool	kredo_get_taxonomy	Browse valid domains/skills

KredoCallbackHandler

Tracks chain execution and builds attestation evidence:

from langchain_kredo import KredoCallbackHandler

handler = KredoCallbackHandler()
chain.invoke(input, config={"callbacks": [handler]})

for record in handler.get_records():
    print(record.build_evidence_context())
    # Chain abc-123
    # Duration: 1500ms
    # Tools used: 3
    # Success rate: 100%
    # ...

    # Use evidence in an attestation
    client.attest_skill(
        subject_pubkey="ed25519:...",
        domain="security",
        skill="incident_response",
        proficiency=3,
        context=record.build_evidence_context(),
        artifacts=record.build_artifacts(),
    )

The handler never submits automatically. You decide when and what to attest.

KredoTrustGate

Policy enforcement for agent selection:

from langchain_kredo import KredoTrustGate

gate = KredoTrustGate(client, min_score=0.3, block_warned=True)

# Non-throwing check
result = gate.check("ed25519:...")
if result.passed:
    ...

# Throwing enforcement
result = gate.enforce("ed25519:...")  # raises InsufficientTrustError

# Decorator
@gate.require(min_score=0.7)
def sensitive_operation(pubkey: str):
    ...

# Select best candidate
best = gate.select_best(
    ["ed25519:agent-a", "ed25519:agent-b"],
    domain="security",
)

Design Decisions

• No auto-attestation — The callback handler collects evidence but never submits. Attestations are cryptographic claims with reputation consequences.
• attest_warning not a tool — Behavioral warnings are too serious for LLM autonomy. Available on the client but not exposed as a LangChain tool.
• No local store — SDK uses the Discovery API only. LangChain developers query api.aikredo.com, not a local database.
• Sync with async passthrough — Kredo uses stdlib urllib (sync). Async tool methods delegate to sync. Acceptable for v0.1.

Development

cd langchain-kredo
pip install -e ".[dev]"
pytest tests/ -v

License

MIT