pydantic-ai-glen

A Pydantic AI capability that connects an agent to Glen, a shared team memory. It does two things:

• Search: adds a glen_search tool so the agent can ask Glen a question.
• Recording: after every run, sends that run's conversation to Glen so it becomes searchable memory. Glen extracts memory in the background on its side.

Install

pip install pydantic-ai-glen

Requires Python 3.10+ and pydantic-ai-slim>=2.9.

Setup

1. Create an API key at https://app.tryglen.com/settings/api-keys.
2. In Glen, switch to the organization the agent should use, click Connect an agent, then MCP server, and copy the value next to x-glen-org-id.
3. Set both where the app runs (never commit them):

GLEN_API_KEY=<your key>
GLEN_ORG_ID=<your organization id>

GLEN_URL is optional and defaults to https://app.tryglen.com.

Usage

from pydantic_ai import Agent
from pydantic_ai_glen import Glen

glen = Glen()  # reads GLEN_API_KEY / GLEN_ORG_ID
agent = Agent("openai:gpt-5", capabilities=[glen])

result = await agent.run("why does the billing worker batch at 500?")
await glen.flush()  # before a short-lived process exits

Recording hooks into the run lifecycle, so run, run_stream, iter and run_sync are all covered with no wrapper code.

Conversations are grouped by Pydantic AI's conversation_id. Pass your own thread id to agent.run(..., conversation_id=...) to keep one Glen conversation across runs.

Compared with the Claude Code plugin

The Claude Code plugin works through hooks. This is what carries over to a Pydantic AI agent:

The plugin also shows a banner when Glen thinks a saved skill fits your prompt (skillsToOffer in each turn response). That only makes sense with a person reading the reply, so a library agent does not get it; it can search skills itself.

Behavior

• A Glen failure never fails a run. Search returns a short message to the model; recording retries transient errors (429/5xx) three times with the same idempotency key, then logs.
• A run that fails or is cancelled is recorded too (what it did so far, ending with a note of how it ended), and its error still raises. A run that fails before its first model request records nothing.
• Only this run's new messages are sent, with tool calls, tool results and thinking as ordered segments. System prompts and non-text prompt content (images, audio) are not sent.
• Everything recorded becomes real team memory. Don't point test runs at your real org.

Known issue

Conversations recorded this way currently show up in Glen labelled as Claude Code sessions (claude-code:<conversationId>). Only the label is affected; recording, search and memory work normally.

Try it

examples/demo.py is an interactive chat that searches Glen and records the conversation:

export GLEN_API_KEY=... GLEN_ORG_ID=...
uv run --with "pydantic-ai-slim[anthropic]" python examples/demo.py --model anthropic:claude-sonnet-5-5
uv run --with "pydantic-ai-slim[anthropic]" python examples/demo.py --model anthropic:claude-sonnet-5-5 --search-only

It asks before recording, since whatever you type becomes team memory. Use any model string Pydantic AI supports (--model or DEMO_MODEL).

Development

uv sync --extra dev
uv run ruff check . && uv run ruff format --check .
uv run pytest

Tests fake Glen's HTTP API with httpx.MockTransport; the agent, capability and message conversion run for real.

See CONTRIBUTING.md and SECURITY.md.

This is an independent integration; Pydantic AI and Glen are their respective owners' projects.

License

MIT, see LICENSE.