Erkle MCPサーバー

npx @erkle/cli init

Your agents forget what your team decided. One teammate's Claude Code learns that the billing service must never retry a charge; tomorrow another teammate's Cursor retries it. CLAUDE.md helps inside one repo, but it doesn't travel across repos, tools, or people, and nobody reviews what goes in it. Erkle gives every agent on the team the same reviewed memory.

What it looks like

After init, a session goes like this (illustrative):

You:    Add retries to the charge endpoint.

Agent:  [memory_search "charge endpoint retries" repo=acme/billing]
        Found 1 constraint (team scope, approved by Dana):
        "Never retry POST /charges. Stripe idempotency keys expire after 24h,
         and a retry after that double-charges. Retry only the webhook."
        I'll add retries to the webhook handler instead and leave /charges alone.

        ...

Agent:  I learned the webhook handler must be idempotent on event.id.
        Propose it as a convention for acme/billing? [y/n]
You:    y
Agent:  [memory_create] Proposed. A maintainer reviews it before it goes live.

What is Erkle?

Erkle is shared, reviewed team memory for AI coding agents. It's a hosted MCP server (https://erkle.ai/mcp) where your team's decisions, conventions, constraints and lessons live, scoped to the team, a project, or a repository. Agents search it before a task and propose what they learned after; maintainers review proposals before they go live. One memory works across Claude Code, Cursor, Codex and claude.ai, so every teammate's agent knows what the team knows.

@erkle/cli is the setup step: erkle init wires your agents to that server and tells them how to use it.

Run it in a git repository (installed globally, the command is erkle init). It:

1. Adds the erkle MCP server to each agent you pick (Claude Code, Cursor, Codex).
2. Adds a short instructions block to AGENTS.md / CLAUDE.md, so agents search Erkle before a task and propose what they learned after.
3. Prints the sign-in step for each agent and a first-run prompt that imports what the repo already knows.

init never signs in and never talks to Erkle. Each agent runs its own browser sign-in the first time it connects.

What it writes

Agent	Project (default)	--global
Claude Code	.mcp.json	claude mcp add --scope user (~/.claude.json)
Cursor	.cursor/mcp.json	~/.cursor/mcp.json
Codex	.codex/config.toml (loaded once you trust the project)	~/.codex/config.toml

Project files live in the repo root, so commit them and every teammate gets the same setup.

Instructions go between <!-- erkle:start --> and <!-- erkle:end --> in AGENTS.md and CLAUDE.md, whichever exist. A CLAUDE.md that is a symlink to AGENTS.md, or that imports it with an @AGENTS.md line, is left alone so the block appears once. If neither file exists, init creates AGENTS.md, which Claude Code, Cursor and Codex all read. Re-running replaces only the text between the markers.

Existing files keep everything else: other MCP servers, other keys, JSON indentation, and TOML comments. If a file can't be parsed (JSON with comments, for example) init stops and writes nothing. If an erkle entry already points somewhere else, init reports it as a conflict and leaves it alone unless you pass --force; everything else in the plan is still applied.

init treats the repo as untrusted, since it may be a fresh clone. If a project file is a symlink that resolves outside the repo, init stops and writes nothing. An existing erkle entry with keys that can run code (command, env, headersHelper, ...) is a conflict too. If a file changes between the plan and the write, init stops instead of overwriting it.

Running init again when everything is set up changes nothing.

Options

--agents <list>     claude,cursor,codex (default: ask, or the detected ones)
--global            write user-level config instead of the repo's
--url <url>         MCP server URL (default https://erkle.ai/mcp, or $ERKLE_URL)
-y, --yes           apply without asking
--dry-run           print the plan and write nothing
--force             replace an existing erkle entry that has a different URL
--no-instructions   don't touch AGENTS.md / CLAUDE.md
-h, --help          show help
-v, --version       show the version

Exit codes: 0 success, nothing to do, or cancelled at the prompt; 1 error (nothing written); 2 bad usage. When stdin or stdout isn't a terminal, init won't ask, so pass --yes to apply (or --dry-run to preview); without either it exits 2.

For local development against a dev server:

ERKLE_URL=http://localhost:8080/mcp npx @erkle/cli init

FAQ

Does init need an Erkle account or API key? No. It only writes config files. Each agent opens a browser OAuth sign-in the first time it connects. There are no API keys.

Which agents are supported? Claude Code, Cursor and Codex. claude.ai connects to the same server as a custom connector; see erkle.ai.

How is this different from CLAUDE.md or AGENTS.md? Instruction files live in one repo and change by commit. Erkle is shared across repos and tools, scoped by team, project and repo, permissioned, and reviewed. init keeps a short block in your instruction files that points agents at Erkle; see Erkle vs. AGENTS.md.

What does it cost? $5 per user per month, or $50 per user per year, after a 7-day trial with no card needed. See pricing.

Who maintains this

Erkle is built and operated by Justin Albrecht. Development happens in a private repository. This public repository is the release mirror: each release lands here as one commit made by the release automation, which is why commits show github-actions[bot], and its tag is what gets published to npm. How Erkle itself handles your data is on the security page.

Bugs

Open an issue at github.com/erkle-ai/cli. This repository receives one commit per release and doesn't take pull requests; see CONTRIBUTING. Report vulnerabilities privately, as described in SECURITY.