Claude Code, Codex & Cursor

The Jackdaw plugin connects Claude Code, Codex and Cursor to your workspace with one install and one sign-in. It does three things:

Context on every prompt. Before your agent answers, Jackdaw looks up what is relevant to the prompt (decisions, rules, traps, the state of the work, the people and projects involved) and hands it to the agent. On Cursor, which can't add context to a prompt, the agent asks for it with Jackdaw's tools.
Session memory. After each finished turn, Jackdaw keeps that turn as dated work memory: what was decided and why, rules you gave the agent, traps and their fixes, what was committed, merged or deployed, and what was left open. Your next session (and your other tools) start with it.
One memory instead of two. Setup offers to import the repository's existing Claude Code memory notes and to turn Claude Code's own memory off, so there is one memory to keep current.

Set it up

Download the plugin from Downloads and unzip it.
From the repository you work in, run:
sh ~/Downloads/jackdaw-agent-plugin/install.sh
It installs the plugin into Claude Code and/or Codex, whichever you have.
Sign in once and choose the workspace this code belongs to. Answer the two questions about importing your notes and turning Claude Code's memory off.

Start a new session and ask about your work ("why do we retry webhooks from a queue?"). The answer should come from your Jackdaw context.

What leaves your computer

Each prompt, to look up the relevant context.
After each turn, your prompt and the agent's reply. Tool calls and their output, the agent's private reasoning, and text the tools inject into a prompt (system reminders, CLAUDE.md / AGENTS.md) are never sent. Passwords, keys and tokens are replaced with [redacted secret] on your computer, and again on the server.

What comes from your sessions is visible only to you. In a shared workspace, teammates can see that your agent read the workspace's memory and whether it found something, never what you asked.

Commands

python3 ~/.local/share/jackdaw-agent-plugin/plugins/jackdaw/scripts/jackdaw-cli capture status   # on or off, what waits to send
python3 ~/.local/share/jackdaw-agent-plugin/plugins/jackdaw/scripts/jackdaw-cli capture off      # stop sending turns
python3 ~/.local/share/jackdaw-agent-plugin/plugins/jackdaw/scripts/jackdaw-cli import-memory --dry-run
python3 ~/.local/share/jackdaw-agent-plugin/plugins/jackdaw/scripts/jackdaw-cli capture-history --days 14 --dry-run

import-memory sends the repository's memory notes; capture-history sends its past sessions. Both say what they would send with --dry-run first. What an imported note or past session says is remembered as of the day it was written, so a note that has gone out of date reads as history, not as the current state.

To disconnect, run jackdaw-cli disconnect, or revoke the connection in Jackdaw.

Tools the agent can call

Besides the context it gets on every prompt, the agent can call Jackdaw directly: get_context for a deeper look, read_evidence to open the original a fact came from, and query for exact counts and lists (for example open merge requests, if your GitLab or GitHub connection imports them). See Connect via MCP.

Optional: the "keep going" agent

A separate, older integration lets a Jackdaw agent decide whether Claude Code should keep going when it tries to stop ("don't stop until the tests pass"). It needs an agent triggered on Claude Code events and a hook snippet from the Agents view, which carries a personal token: anyone with it can post to your account, so regenerate it in the same panel if it leaks. You don't need it for context or session memory; the plugin covers both.