Langclaw CROO Agent

Langclaw CROO Agent is a paid callable research agent for CROO CAP and MCP-compatible coding tools.

It turns a research request into a concise report with sources, confidence, provider trace, and delivery proof. The same core agent runs behind a CROO provider, a hosted MCP server, a local MCP server, and a Codex plugin.

Langclaw can run without a model API key for non-onchain services. Langclaw Onchain Intelligence can optionally use a semantic reasoning layer on the operator VPS, then validate and run Langclaw's read-only onchain tools before returning delivery proof.

What It Does

• Accepts paid research orders through CROO CAP.
• Exposes Langclaw Onchain Intelligence as a 0.10 USDC service for requester agents.
• Exposes MCP tools for Codex, Claude Code CLI, Cursor, Windsurf, and other MCP clients.
• Ships a Codex plugin with a small skill that tells an agent when to call Langclaw.
• Produces source-backed research, onchain intelligence, claim checks, builder submission reviews, and readiness checks.

Main Tools

• langclaw_research: Research a protocol, project, token, market topic, or agent idea.
• langclaw_verify_claim: Check whether a claim has enough source support.
• langclaw_builder_review: Review whether a project fits a builder program, hackathon, grant, or submission goal.
• langclaw_onchain_intelligence: Run read-only chain, token, wallet, contract, transaction, bridge, governance, or protocol analytics.
• langclaw_readiness: Check local provider, model, MCP, and CROO configuration.

Strong onchain prompts sent through langclaw_research also auto-route to onchain intelligence. Examples include wallet, contract, transaction, TVL, liquidity, pool, bridge, governance, token holder, and chain activity requests.

User Install

Users do not need to clone this repository or bring provider API keys. The hosted Langclaw server runs on the operator VPS and keeps the provider keys there.

Codex one-line install:

npx @langclaw/mcp-client install-codex --url https://langclaw.nanta.tech/mcp --token LICENSE_TOKEN

This installs the hosted MCP config, Langclaw skill, and Langclaw plugin.

Codex also understands slash-style Langclaw prompts through the installed skill:

/langclaw-research CROO agent market fit
/langclaw-onchain Base smart money last 7 days
/langclaw-builder-review Review this repo for CROO
/langclaw-verify Langclaw supports hosted MCP access
/langclaw-readiness

Claude Code CLI one-line install:

npx @langclaw/mcp-client install-claude --url https://langclaw.nanta.tech/mcp --token LICENSE_TOKEN

Cursor one-line install:

npx @langclaw/mcp-client install-cursor --url https://langclaw.nanta.tech/mcp --token LICENSE_TOKEN

Windsurf one-line install:

npx @langclaw/mcp-client install-windsurf --url https://langclaw.nanta.tech/mcp --token LICENSE_TOKEN

Install without plugin:

npx @langclaw/mcp-client install-codex --url https://langclaw.nanta.tech/mcp --token LICENSE_TOKEN --no-plugin

Uninstall from Codex:

npx @langclaw/mcp-client uninstall-codex

Manual MCP client config:

{
  "mcpServers": {
    "langclaw": {
      "command": "npx",
      "args": [
        "@langclaw/mcp-client",
        "--url",
        "https://langclaw.nanta.tech/mcp",
        "--token",
        "LICENSE_TOKEN"
      ]
    }
  }
}

The same manual config shape works for Codex, Claude Code CLI, Cursor, Windsurf, and MCP clients that launch stdio servers. Codex remains the richest install because it also writes the Langclaw skill and plugin.

CROO Store

End users can also use Langclaw without installing anything:

1. Open the CROO Agent Store.
2. Select Langclaw Intel.
3. Order Langclaw Builder Pass License for 30-day hosted MCP access.
4. Use the returned license token with Codex, Claude Code CLI, Cursor, Windsurf, or another MCP client.
5. Order langclaw.research.brief or langclaw.onchain.intelligence for direct CROO delivery.

Recommended CROO dashboard service:

Service Name: Langclaw Onchain Intelligence
Description: Source-backed onchain intelligence for agents that need read-only chain, token, wallet, contract, protocol, and market context. Langclaw plans the request, runs validated onchain tools, and returns a structured intelligence packet with sources, risks, opportunities, and reusable agent context.
Price: 0.10 USDC
SLA: 15 minutes
Deliverable: Text JSON
Requirements: Schema

Store listing evidence to include in the final submission:

• Public CROO Agent Store listing URL.
• Screenshot that shows service name, price, input fields, output format, and SLA.
• Screenshot or export that shows the service is discoverable by humans or requester agents.
• Service ID for Langclaw Onchain Intelligence.

VPS Operator Setup

npm install
cp .env.example .env

Store provider keys only on the VPS:

• TAVILY_API_KEY
• BRAVE_SEARCH_API_KEY
• GITHUB_TOKEN
• ETHERSCAN_API_KEY
• ALCHEMY_API_KEY
• GOPLUS_API_KEY
• GOPLUS_API_SECRET
• DUNE_API_KEY
• DUNE_API_KEYS
• DUNE_API_KEY_RETRY_STATUS
• DUNE_API_KEY_MAX_ATTEMPTS
• OPENAI_API_KEY
• CROO_API_KEY
• CROO_TARGET_SERVICE_ID
• LANGCLAW_ONCHAIN_SERVICE_ID
• LANGCLAW_ACCESS_TOKENS
• LANGCLAW_ADMIN_ACCESS_TOKENS
• LANGCLAW_LICENSE_STORE_PATH
• LANGCLAW_LICENSE_DEFAULT_DAYS
• LANGCLAW_LICENSE_DEFAULT_CALLS
• LANGCLAW_CROO_EVIDENCE_LOG_PATH
• LANGCLAW_A2A_WORKBENCH_ENABLED
• LANGCLAW_A2A_WORKBENCH_REQUIRED
• LANGCLAW_A2A_WORKBENCH_SERVICE_ID
• LANGCLAW_A2A_WORKBENCH_CAPABILITY_ID
• LANGCLAW_A2A_WORKBENCH_TIMEOUT_MS
• LANGCLAW_A2A_WORKBENCH_MAX_INPUT_CHARS

Start the hosted MCP server:

npm run server

Start the CROO provider separately:

npm run croo:provider

Local Development MCP

npm run mcp

Example MCP client command:

{
  "mcpServers": {
    "langclaw-croo-agent": {
      "command": "npm",
      "args": ["--prefix", "/Users/ridhorinanta/hackathon/langclaw-croo-agent", "run", "mcp"]
    }
  }
}

Local Client Setup

Codex plugin:

• Plugin manifest: plugin/.codex-plugin/plugin.json
• Plugin MCP config: plugin/.mcp.json
• Plugin skill: plugin/skills/langclaw-croo-agent/SKILL.md

Codex MCP:

• Add the MCP command above to Codex MCP settings.
• Keep this repo path stable so the stdio command can find the server.

Claude Code CLI:

• Add an MCP server named langclaw-croo-agent.
• Use npm as the command.
• Use --prefix, this repo path, run, and mcp as args.

Cursor:

• Add the same stdio MCP server command to Cursor MCP settings.
• Restart Cursor after adding the server.

Windsurf:

• Add the same stdio MCP server command to Windsurf MCP settings.
• Restart the MCP server if .env changes.

Run CROO Provider

npm run croo:provider

Required live settings:

• CROO_API_URL
• CROO_WS_URL
• CROO_SDK_KEY or CROO_API_KEY
• LANGCLAW_PROVIDER_FUND_ADDRESS for CROO fund-transfer services
• LANGCLAW_AGENT_PRICE_USDC
• LANGCLAW_ONCHAIN_SERVICE_ID for the CROO dashboard service ID of Langclaw Onchain Intelligence
• LANGCLAW_ACCESS_TOKENS for hosted MCP access control
• LANGCLAW_ADMIN_ACCESS_TOKENS for hosted MCP admin fallback access
• LANGCLAW_LICENSE_STORE_PATH for paid license token storage
• LANGCLAW_CROO_EVIDENCE_LOG_PATH for CAP order evidence logs

CROO_API_KEY is the name shown by the CROO dashboard. CROO_SDK_KEY is also supported for compatibility with SDK docs. Do not export a private key from your main browser wallet for this repo. CROO shows the agent account wallet in its dashboard. Fund that agent account when CROO requires balance for paid actions.

LANGCLAW_PROVIDER_FUND_ADDRESS must be a Base mainnet EVM address such as 0x.... The provider only uses it when a CROO negotiation includes fund-transfer fields. Non-fund services keep the normal CROO escrow flow.

The provider appends safe CAP lifecycle evidence to LANGCLAW_CROO_EVIDENCE_LOG_PATH. Each JSONL row records the stage, negotiation ID, order ID, capability, input hash, settlement mode, delivery hash, and source count when available. It does not store raw buyer prompts, API keys, or license tokens.

For A2A smoke tests, set CROO_TARGET_SERVICE_ID to the same service ID as LANGCLAW_ONCHAIN_SERVICE_ID. Requester agents should send langclaw.onchain.intelligence requirements with research_prompt, chain, scope, timeframe, targetUse, and responseLanguage. query remains supported as a backward-compatible alias.

Langclaw Onchain Intelligence can optionally hire Universal Workbench as a downstream A2A agent. Enable it with LANGCLAW_A2A_WORKBENCH_ENABLED=true. Langclaw uses its own CROO agent key as the requester, so buyers do not provide a Langclaw SDK key. The default downstream target is service a8f1c20d-73f4-4551-856a-32315e18d261 with capability universal.workbench.agent. Keep LANGCLAW_A2A_WORKBENCH_REQUIRED=false for live reliability. If Workbench times out, Langclaw still delivers the onchain intelligence packet and omits a2aWorkPack from the buyer-facing delivery.

License Tokens

Paid MCP access uses a license token. The default paid license lasts 30 days and allows 300 MCP tool calls.

Create a demo license:

npm run license:create -- --label demo --days 7 --calls 50

List licenses:

npm run license:list

Revoke a license:

npm run license:revoke -- --token lc_live_xxx

The license registry stores token hashes only. Keep data/licenses.json on the VPS and do not commit it.

Optional onchain settings:

• ETHERSCAN_API_KEY
• ALCHEMY_API_KEY
• GOPLUS_API_KEY
• GOPLUS_API_SECRET
• DUNE_API_KEY
• DUNE_API_KEYS
• DUNE_API_KEY_RETRY_STATUS
• DUNE_API_KEY_MAX_ATTEMPTS
• DUNE_DEFAULT_QUERY_ID
• OPENAI_API_KEY
• OPENAI_MODEL
• LANGCLAW_ONCHAIN_REASONING_ENABLED
• LANGCLAW_ONCHAIN_REASONING_EFFORT
• LANGCLAW_ONCHAIN_REASONING_REQUIRED
• LANGCLAW_ONCHAIN_REASONING_MAX_INPUT_CHARS
• LANGCLAW_ONCHAIN_REASONING_MAX_OUTPUT_TOKENS
• LANGCLAW_ONCHAIN_REASONING_MAX_RETRIES
• LANGCLAW_A2A_WORKBENCH_ENABLED
• LANGCLAW_A2A_WORKBENCH_REQUIRED
• LANGCLAW_A2A_WORKBENCH_SERVICE_ID
• LANGCLAW_A2A_WORKBENCH_CAPABILITY_ID
• LANGCLAW_A2A_WORKBENCH_TIMEOUT_MS
• LANGCLAW_A2A_WORKBENCH_MAX_INPUT_CHARS

Onchain semantic reasoning is scoped only to Langclaw Onchain Intelligence. It receives the buyer research_prompt and safe provider summaries, not raw provider payloads, env dumps, API keys, private keys, license tokens, or full evidence logs. Keep LANGCLAW_ONCHAIN_REASONING_REQUIRED=false for demo fallback unless you want paid onchain orders to fail when the semantic layer is unavailable.

Use LANGCLAW_PROVIDER_MODE=mock to run a local dry run without opening a live provider connection.

Check live credential readiness:

npm run smoke:croo-live

Run a live requester smoke order from a separate CROO requester agent:

npm run croo:requester-smoke
npm run croo:evidence-report

The requester smoke uses the requester agent's own CROO key. It reads CROO_REQUESTER_SDK_KEY first for shared local .env files, then falls back to the official CROO_SDK_KEY name from the CROO docs. Do not share the Langclaw provider key with buyers or external requester agents.

The requester smoke writes a redacted summary to data/croo-requester-smoke.json. The report command turns provider and requester evidence into docs/CROO_LIVE_EVIDENCE.md for demo and submission review.

The data/ directory is local runtime output and is intentionally ignored. Keep shareable submission evidence in docs/CROO_LIVE_EVIDENCE.md.

For final commerce proof, capture at least one paid requester order with requester agent ID, requester wallet, service ID, negotiation ID, order ID, payment transaction hash, delivery transaction hash when available, and delivery hash.

Development

npm run typecheck
npm test
npm run build
npm run smoke:readiness
npm run smoke:croo-live
npm run validate:package
npm run scan:cleanup
npm run scan:dash

Submission Story

Langclaw CROO Agent shows agent commerce in practice. A buyer can pay for a callable research or onchain intelligence capability, receive a verifiable delivery payload, and reuse the same hosted intelligence from coding tools through remote MCP without cloning this repository.

Anti-Sybil Proof

Use real requester agents and buyer wallets for final proof. Keep provider and requester keys separate. Do not use fake payments, self-trade loops, or synthetic order activity as reward evidence. Keep redacted CROO order logs, requester smoke summaries, transaction hashes, and Store listing screenshots available for human review.

Submission Docs

• Demo script
• DoraHacks submission copy
• Sample payloads