Cognee Connection And Status

Plugin root. The commands below use ${CURSOR_PLUGIN_ROOT}, which Cursor does not set in your shell. Take the path from the Cognee plugin root: line in the session context (or use the directory two levels above this SKILL.md, the one holding scripts/) and run export CURSOR_PLUGIN_ROOT="<that path>" in the same shell command, before the skill command.

Use this skill when the user asks to configure Cognee, check whether Cognee is ready or connected, switch between local and cloud mode, or diagnose a connection problem.

Rules

• Server first. Answer every question here from the plugin's own diagnostic and the running Cognee server over HTTP. That is the authoritative path in both local and cloud mode.
• cognee-cli is a last resort, not an alternative. It is reachable only on a machine holding a cognee source checkout: scripts/cognee-cli.sh exits 64 ("Run this from the Cognee repository root or set COGNEE_REPO_ROOT") anywhere else, and in cloud mode the plugin's venv is never built at all. Never reach for it before the two steps below have actually failed.
• Never run uv sync or any dependency install. On a plugin install there is no cognee repository to sync, and in the user's own project it would rebuild their dependencies.
• Do not print secret values from .env, config files, shell history, or command output. Report whether a key appears configured, never the value.
• If a command may create, delete, or overwrite durable Cognee state, say what it will affect before running it.
• Do not use MCP for this plugin.

1. Check the mode and the connection

One command answers "is Cognee ready?", "which mode am I in?", and "what is wrong?" — it imports no cognee, makes one /health probe, and works from any directory:

python3 "${CURSOR_PLUGIN_ROOT}/scripts/doctor.py" --json

The JSON carries:

Field	Meaning
mode	Local or Cloud, plus an annotation naming what forced the decision (e.g. forced by COGNEE_BACKEND=local)
server_url	The endpoint this terminal actually resolved
reachable	Whether the health probe succeeded
latency_ms	Round-trip time, or null when unreachable
api_key_source	Which credential the plugin is using (plugin identity, env var, cached, minted)
memory_sharing	Shared-agent-memory state
circuit_breaker	Whether recall is currently short-circuited

Read mode before doing anything mode-specific. In particular: in Cloud mode there is no local server to start and no local UI to launch — see the local-ui skill, which stops on cloud by design.

Drop --json for a human-readable table when reporting to the user.

2. Probe the server directly (optional)

When you need to confirm the endpoint yourself, or the doctor's verdict looks stale:

curl -sS -i "${COGNEE_BASE_URL:-http://localhost:8011}/health"

http://localhost:8011 is the plugin's local server. Do not substitute :8000 — that is the cognee-cli's own default for a server it would start itself, unrelated to the one the plugin runs and talks to.

For an authenticated check of the resolved credential (cloud, or any server with auth enabled), use the forget helper's credential resolver rather than guessing where the key lives:

eval "$(${CURSOR_PLUGIN_ROOT}/scripts/cognee-forget.sh env)" && \
curl -sS -o /dev/null -w '%{http_code}\n' \
  -H "X-Api-Key: ${COGNEE_API_KEY}" "${COGNEE_BASE_URL}/api/v1/users/me"

200 means the key is accepted; 401/403 means it was rejected by that server. In local mode the key is never exported to the shell, so a raw request without this resolver will 401 misleadingly.

Configure the connection

Durable configuration (COGNEE_BASE_URL, COGNEE_API_KEY, LLM_API_KEY, ...) belongs in ~/.cognee/.env — a one-time setup file shared with the Claude Code plugin, loaded at session start. Guide the user to edit that file rather than exporting in every shell; never echo its secret values. Shell exports still override it per terminal. Changes apply on the next cursor launch.

Cloud or a remote server — both values:

mkdir -p ~/.cognee
cat >> ~/.cognee/.env <<'EOF'
COGNEE_BASE_URL="https://your-instance.cognee.ai"
COGNEE_API_KEY="ck_..."
EOF
chmod 600 ~/.cognee/.env

Local (default when COGNEE_BASE_URL is unset) — the plugin bootstraps a local Cognee API on http://localhost:8011 and auto-mints COGNEE_API_KEY, so only the LLM key is required:

mkdir -p ~/.cognee
cat >> ~/.cognee/.env <<'EOF'
LLM_API_KEY="sk-..."
EOF
chmod 600 ~/.cognee/.env

Re-pasting either block is safe — the last value of a repeated key wins.

Which mode wins

The file may hold both modes' variables at once. The mode is then decided per terminal, in this order:

1. An exported COGNEE_BACKEND (or the plugin-specific COGNEE_CURSOR_BACKEND, which beats it) pins that terminal: local or cloud.
2. Otherwise cloud wins when COGNEE_BASE_URL is set anywhere.
3. Otherwise local.

A forced mode is pinned: COGNEE_BACKEND=cloud with no URL configured stays cloud and fails visibly (✕ (missing_cognee_base_url)) rather than silently falling back. unset COGNEE_BASE_URL is not a way to go local — the env file re-injects it on the next launch; use the switch.

COGNEE_BACKEND may also live in ~/.cognee/.env to make a mode the durable default. The doctor's mode field always reports what the terminal actually resolved, with the cause.

Fallback — cognee-cli, local mode with a cognee checkout only

Only if steps 1 and 2 both failed, the mode is Local, and the machine has a cognee source checkout. Route it through the wrapper so a missing checkout fails with a clear message instead of an opaque uv error:

"${CURSOR_PLUGIN_ROOT}/scripts/cognee-cli.sh" config list

Set COGNEE_REPO_ROOT to the checkout path if the wrapper cannot find it. Use config get <KEY> only for non-secret settings.

Exit code 64 means there is no cognee checkout here — that is expected on a normal install, and it is not a Cognee fault to report. Say the server is unreachable and give the user the doctor output instead.