# Erkle

> Erkle is shared, governed team memory for AI coding agents. It is a hosted MCP server at https://erkle.ai/mcp that gives Claude Code, Cursor, Codex, and claude.ai one reviewed, permissioned memory shared across a team's people, repositories, and tools.

Key facts:

- Connect: `npx @erkle/cli init` in a repo, or one MCP URL (`https://erkle.ai/mcp`, Streamable HTTP) and browser OAuth sign-in. No API keys.
- Agents search Erkle at the start of a task and propose what they learned after. Search is keyword plus semantic, returns at most 8 results, and stays within about 4,000 tokens.
- Scopes: every memory belongs to one scope: the team, a project, or a repository. Projects group repositories, and a repository can belong to several projects. A search from a repository reads that repository, every project it belongs to, and the team.
- Review: maintainer writes go live; contributor writes arrive as proposals. Edits to an approved memory wait as a pending revision while the approved version stays in use.
- Versions: every edit is a new immutable version. A concurrent edit is rejected with the current version instead of overwriting it.
- Permissions: memories a person can't read don't appear at all, not even as "access denied". Removing a member or revoking an agent applies on the next request.
- Secret guard: writes containing private keys or AWS, GitHub, Slack, OpenAI, Anthropic, or Stripe tokens are rejected.
- Memory types: decision, constraint, convention, how-to, lesson, interface contract, plan, strategy. A memory has a title, a body of up to 8,000 characters, tags, a source URL, an author, and the agent that wrote it.
- Import: an agent can import a repository's CLAUDE.md, docs, and ADRs, up to 100 memories per call, with a result for each.

Pricing:

- One plan: $5 per user per month, or $50 per user per year (two months free). Every member of a team is billed, on one invoice per team. Every feature is included. No metering on searches or memories; rate limits exist only to stop abuse.
- Trial: a person's first team gets a 7-day trial. Signing up needs only a verified email, no card. A team that subscribes during its trial keeps the remaining days and is first charged when the trial ends. One trial per person.
- Members: on annual billing, an added member is charged the prorated rest of the year at once; on monthly billing, the prorated amount goes on the next invoice. Removed members come off the bill at the next renewal, with no refund.
- Unpaid teams: only MCP (agent) access pauses until the team pays. The dashboard, review, history, members, and projects keep working.
- Cancelling keeps access until the end of the paid period. Deleting a team cancels its subscription at once, with no refund.
- Self-hosting on your own infrastructure is priced separately. Contact hello@erkle.ai.

Company:

- Erkle is built and operated by Justin Albrecht (https://www.linkedin.com/in/justin-albrecht-19a021bb/). Contact: hello@erkle.ai.
- Privacy: memories never leave Erkle's infrastructure, and no AI provider receives them. Search embeddings are computed by Erkle's own embedding service on a separate Erkle server. A team's own agents receive the memories they search for, under that team's permissions.
- Hosted on DigitalOcean in San Francisco, USA. The database, swap, and backups are encrypted at rest, and traffic between Erkle's servers goes over an encrypted WireGuard tunnel inside a private network. Resend (email), Stripe (billing), and Google and GitHub (optional sign-in) handle account data, never memory content. No SOC 2 or ISO 27001 certification yet; security questionnaires are answered on request. Details: https://erkle.ai/security

## Pages

- [Home](https://erkle.ai/): what Erkle does, how capture, review, and reuse work, scopes, and trust features
- [Pricing](https://erkle.ai/pricing): the plan, monthly and annual billing compared, and billing FAQ
- [Compare](https://erkle.ai/compare): how Erkle relates to instruction files and built-in agent memory, and when to use each
- [Erkle vs. CLAUDE.md](https://erkle.ai/compare/claude-md): when Claude Code's instruction file is enough, and when to add shared memory
- [Erkle vs. AGENTS.md](https://erkle.ai/compare/agents-md): the open instruction-file format compared with a scoped, reviewed team memory
- [Erkle vs. built-in agent memory](https://erkle.ai/compare/built-in-memory): personal, per-tool memory compared with shared team memory
- [Security](https://erkle.ai/security): hosting, encryption, access control, agent sign-in, the secret guard, subprocessors, backups, certifications, and vulnerability reporting
- [Privacy policy](https://erkle.ai/privacy): what Erkle collects, which providers handle it, and how to delete it. Erkle never sells data.
- [Terms of service](https://erkle.ai/terms): plain-English terms for using Erkle

## Connect an agent

- Any repo, all agents at once: run `npx @erkle/cli init`. It adds the `erkle` MCP server to Claude Code (`.mcp.json`), Cursor (`.cursor/mcp.json`), and Codex (`.codex/config.toml`), and adds an instructions block to `AGENTS.md` / `CLAUDE.md`. It only writes config; each agent signs in with OAuth on first connect. Source: https://github.com/erkle-ai/cli
- Claude Code: `claude mcp add --transport http erkle https://erkle.ai/mcp`, then run `/mcp`, pick `erkle`, and choose Authenticate.
- Cursor: add `"erkle": { "url": "https://erkle.ai/mcp" }` under `mcpServers` in `~/.cursor/mcp.json`, then click Connect in Cursor Settings → MCP.
- Codex: add `[mcp_servers.erkle]` with `url = "https://erkle.ai/mcp"` to `~/.codex/config.toml`.
- claude.ai: add a custom connector with the URL `https://erkle.ai/mcp`.

## Optional

- [Contact](mailto:hello@erkle.ai): sales and self-hosting