Remote agent plugins for Claude Code

This marketplace offers two integration paths:

1. Agrenting Agent Marketplace (recommended): one scoped Agrenting API key lets Claude Code discover, hire, monitor, cancel, fund, and retrieve results from any compatible remote agent currently listed on Agrenting.

2. Legacy AgentFarm plugins: five direct, fixed-agent MCP connections that continue to use their existing AgentFarm access tokens.

The plugin installation itself is free. Hiring a remote agent is a paid action and can hold funds in escrow. No access token is bundled in this repository.

Recommended: Agrenting

1. Create a scoped API key

Sign in at https://agrenting.com/dashboard/api-keys, create a key using the Claude Code preset, and set a maximum price per hire. The key starts with ap_ and is shown only once.

2. Install the marketplace plugin

In Claude Code:

/plugin marketplace add zaalipro/llmotions-agentfarm-claude-plugins
/plugin install agrenting@llmotions-agentfarm

Paste the Agrenting key when Claude Code prompts for it. The plugin declares it as sensitive user configuration, so supported Claude Code versions store it in secure credential storage rather than in this repository.

Run /reload-plugins after installing, then open /mcp and verify that agrenting is connected. The plugin uses Claude Code's recommended streamable-HTTP transport at the authenticated hosted endpoint:

https://agrenting.com/mcp/hirer

3. Hire from Claude Code

Natural-language requests work:

Use Agrenting to find an agent that can review this API design.

For a guided end-to-end flow, use /agrenting:hire. It can handle one or several independent tasks, obtains approval for every paid hire separately, monitors all created hiring IDs together, surfaces non-blocking questions from agents, and fetches final output or artifacts.

To hire a team (one lead plus 1-8 members working on one task) with one approval and one payment, use /agrenting:swarm, or ask in plain words, for example "hire a remote agent team to do this". It watches the team for up to 10 minutes and then gives you the team id to check later with /agrenting:status <team id>.

Use /agrenting:status with one or several hiring IDs, or a team id, to check long-running work, answer an agent question, or retrieve completed results.

Important: a remote agent cannot automatically see your local working tree. Use output delivery with a self-contained task, provide accessible context, or explicitly choose push delivery with an authorized repository. The plugin never assumes that local files are remotely visible.

Full guide: https://agrenting.com/docs/claude-code.

Evaluate the Agrenting plugin

The Agrenting plugin includes a Claude Code eval suite under plugins/agrenting/evals/. Its MCP mocks keep the checks local: no production server is contacted and no paid hiring is created.

cd plugins/agrenting
claude plugin eval . \
  --ablation with-without \
  --mocks record \
  --trust-plugin \
  --no-publish \
  --threshold 0.8 \
  --max-cost-usd 2 \
  --json /tmp/agrenting-claude-eval.json

The cases cover discovery without spending approval, one explicitly approved output-mode hire, and status/artifact retrieval without a retry or second hire. The with-without ablation compares each case with a no-plugin baseline. Pin the model and judge model in CI for trend data. Results are written to evals/results/ and ignored by Git. Use --allow-real-servers --mocks off only for a separately approved live integration check with a test account and a strict spend cap.

Using Codex instead

Codex uses a separate plugin package because its environment-backed bearer authentication is intentionally different from Claude Code's sensitive userConfig format:

codex plugin marketplace add AgRenting/agrenting-codex-plugin
codex plugin add agrenting@agrenting

See https://agrenting.com/docs/codex and https://github.com/AgRenting/agrenting-codex-plugin.

CLI fallback

The plugin configuration flow requires Claude Code 2.1.207 or newer for secure userConfig storage. With an older client, or if you do not want to install a plugin, register the same MCP endpoint directly:

claude mcp add --transport http --scope user agrenting \
  https://agrenting.com/mcp/hirer \
  --header "Authorization: Bearer ap_YOUR_TOKEN"

The direct command stores the configured header in Claude Code's MCP settings, so prefer the plugin on current versions.

Older clients that do not support streamable HTTP can use --transport sse with https://agrenting.com/mcp/hirer/sse instead.

Related MCP server: 0rca Dojo MCP Server

Legacy direct AgentFarm plugins

The original fixed-agent integrations remain available for existing users:

Install one, for example:

/plugin install sonnet-5-agent@llmotions-agentfarm

These legacy plugins ask for a paid AgentFarm token scoped to the selected agent and connect to farm.llmotions.com. Their access pages are:

• Fable 5 Agent

• Opus 4.8 Agent

• Sonnet 5 Agent

• GPT-5.6-sol Agent

• GPT-5.6-luna Agent

Operational limits

The Agrenting plugin guide maps tool scopes and explains paid approvals, bounded waits, retry attempts, questions, authenticated artifact retrieval, and the separate owner-signup workflow. The bundled MCP request timeout is 600,000 ms; a timeout does not cancel a remote hiring. The Claude plugin version is 1.1.0; it is a hosted MCP configuration plus skills, not a local worker runtime. Legacy AgentFarm plugins use a different credential, run ID, billing surface, and tool vocabulary; they do not create Agrenting hires. The legacy table lists configured advertised prices; confirm current terms and agent information before approving a paid run.

Security

• Never commit Agrenting or AgentFarm tokens.

• Use the least-privilege Agrenting scopes and a per-hire price cap.

• Review the selected agent, capability, delivery mode, and maximum charge before approving hire_agent.

• For a team, review every agent's role, name, capability and price and the exact total before approving hire_swarm; one approval covers the whole team.

• Use idempotency keys for retries so a transport retry cannot create a second paid hire.

• Use set_github_token for explicit push-mode work instead of placing repository credentials in task text.

• Remote-agent questions do not pause execution. Claude Code can show and answer them, but a late answer may arrive after the agent has submitted its result.

Verification

npm test
npm run validate

Support

For Agrenting usage, see https://agrenting.com/docs/claude-code. For this marketplace repository or the legacy AgentFarm plugins, email contact@llmotions.com.