Skip to content
Erold Docs

Quickstart

Persistent project memory for Claude Code. Captures decisions, bugs, deploys across sessions so any new agent picks up exactly where the last one stopped.

Prerequisites (one-time)

Section titled “Prerequisites (one-time)”

You’ll need:

Verify:

Terminal window
node --version && jq --version && command -v secret

Step 1 — Store credentials in Keychain

Section titled “Step 1 — Store credentials in Keychain”
Terminal window
secret set erold.api-key      # paste API key, hit enter
secret set erold.tenant       # paste tenant ID, hit enter

The values never touch disk outside Keychain. They are not in .env, not in .mcp.json, not in shell history.

Step 2 — Install the plugin

Section titled “Step 2 — Install the plugin”
Terminal window
claude plugin install erold-dev/claude-plugin --scope user

The plugin ships an MCP launcher that reads credentials from Keychain at startup — no claude mcp add needed.

Verify:

Terminal window
claude mcp list           # erold-pm: ✓ Connected

Step 3 — Link a project

Section titled “Step 3 — Link a project”

In any repo:

/erold:init

This creates .erold.json at the project root with only a projectId — no secrets, safe to commit.

From here, just work. Hooks capture activity passively; the local outbox ships events to the backend in the background. Latency is <60 ms per hook.

What you get out of the box

Section titled “What you get out of the box”

Every Claude Code session you open in that project:

  • The agent calls get_context() at startup and inherits: open tasks, open + recently-resolved bugs (with fix narratives), recent deploys, every decision ever made, credential references — and gaps flags telling it what was NOT logged so it can ask you.
  • Hooks capture: file edits (paths only — never contents), Bash commands (hashed + service detection), prompt metadata (hash + intent), subagent starts/stops, compaction checkpoints, session lifecycle.
  • Two redaction layers protect you from accidental secret leaks: path denylist drops .env / .ssh/ / *.pem / secrets/ events entirely; pattern blocklist redacts AWS / Scaleway / Stripe / GitHub / Anthropic / OpenAI keys + JWTs + PEM blocks before they leave your machine.

Not passive — needs explicit log()

Section titled “Not passive — needs explicit log()”

Four things the plugin cannot infer and that the agent should record explicitly. See AGENT_PLAYBOOK.md for triggers:

  • Decisions — log(content, type="decision") or decision_log(...).
  • Bug fix narrative — bug_resolve(root_cause, fix_narrative).
  • Deploy outcome — deploy_log(status=succeeded|failed|...).
  • Credential refs — credential_ref_upsert(name, store, lookup_key, …).

If the agent forgets, the next session’s get_context() returns gaps[] with the missing categories — a built-in safety net.

Quick sanity check

Section titled “Quick sanity check”
Terminal window
tail -n 3 ~/.erold/outbox/events.jsonl   # recent events queued locally
tail -n 5 ~/.erold/error.log             # any transmission errors

Next steps

Section titled “Next steps”