withlore

Connect your voice or text agent to Lore. Lore records its traces, tool calls, conversations and call audio, then simulates users against it and grades every conversation.

Documentation: docs.lore-labs.ai

Create a key in Lore under Settings → Projects & keys, then:

export LORE_API_KEY="lore_sk_..."

Voice agents (LiveKit)

pip install "withlore[livekit]"

In your job entrypoint, before session.start():

from withlore.livekit import observe

await observe(ctx, session)
await session.start(agent=agent, room=ctx.room)

Lore records both sides of the call audio, the transcript with per-turn latency, tool calls and LiveKit's spans. When Lore simulates callers against your agent, the same call records into the simulation instead of production. See the LiveKit guide.

Text and AI agents

Install withlore with the extra for your framework, then call withlore.init() at startup, before you create your model client.

pip install "withlore[openai]"

import withlore

withlore.init()

with withlore.conversation() as convo:
    reply = client.chat.completions.create(model="gpt-5-mini", messages=messages)
    convo.record(messages + [reply.choices[0].message])

withlore.init() records traces. convo.record() records the conversation from the messages your framework already keeps: OpenAI and Anthropic message lists, OpenAI Agents SDK results, LangChain and LangGraph messages, and Claude Agent SDK messages.

See Record conversations for chats that span several requests.

Name your agent

Set both of these so Lore can track the agent and compare its versions. If one process runs several agents, see Agents and versions.

export LORE_AGENT_KEY="support-agent"       # Stays the same
export LORE_AGENT_VERSION="2026-10-19"      # Changes with each release

Simulate and evaluate

withlore.Client and the withlore CLI run scenarios, simulations and grading from code or CI. Start with withlore auth login, then withlore catalog --json. See Evaluate your agent.

Bring an existing eval suite as a lore.library.v1 bundle (personas, metrics, scenarios and suite or holdout collections, each with a stable key): withlore library import bundle.json --dry-run, then without --dry-run. Repeating an import changes nothing; withlore library export prints the project's library for diffing. See Import a library.

Privacy

By default Lore receives operational metadata only: models, token counts, timing and status. Prompts, transcripts, tool payloads and audio are recorded only when your code asks for them and your project admits them. See Privacy.

Links

• Documentation
• Release history

Licensed under Apache-2.0.