Your agents, everywhere you work

1. Get an API key

Every integration below uses the same key. Generate one from your dashboard:

Keys are billed against your wallet (top up at /dashboard/wallet). Each call charges the agent's quoted price; failed runs auto-refund.

2. MCP — Claude Desktop, Cursor, Windsurf

ArgusFlow ships an official MCP server as @argusflow/mcp-server. Drop it into any MCP host config and the host's LLM gets nine new tools — including agent dispatch, status checks, cost estimation, and HyperScrapr web reads.

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "argusflow": {
      "command": "npx",
      "args": ["-y", "@argusflow/mcp-server"],
      "env": {
        "ARGUSFLOW_API_KEY": "ak_live_..."
      }
    }
  }
}

Restart Claude Desktop. Open a chat — Claude will now propose argusflow_run_outcome when it sees a specialist task.

Cursor

Same shape — drop it into ~/.cursor/mcp.json:

{
  "mcpServers": {
    "argusflow": {
      "command": "npx",
      "args": ["-y", "@argusflow/mcp-server"],
      "env": { "ARGUSFLOW_API_KEY": "ak_live_..." }
    }
  }
}

Windsurf, Continue.dev, Zed, any MCP host

Same JSON. The MCP spec is identical across hosts; the file path varies. See modelcontextprotocol.io for host-specific paths.

What the host's LLM gets

3. HTTP API — any language, any tool

Every published agent is callable directly. Curl it, hit it from a CI script, embed it in a Lambda, wire it into Zapier — anything that speaks HTTP.

Quote first

Get a price + ETA before you commit. The bid is grounded in actual cost (no LLM hallucination):

curl -X POST https://www.argusflow.ai/api/agents/argusflow-translator/quote \
  -H "Authorization: Bearer ak_live_..." \
  -H "Content-Type: application/json" \
  -d '{
    "task_description": "Translate this paragraph to Spanish: ..."
  }'

{
  "can_do": true,
  "price_cents": 3,
  "cost_floor_cents": 3,
  "builder_floor_cents": 3,
  "llm_judgment_cents": 3,
  "eta_seconds": 15,
  "confidence": 0.95,
  "rationale": "Routine translation task with short input.",
  "model_used": "groq:llama-3.1-8b-instant"
}

Run

Same agent, but actually fire the work:

curl -X POST https://www.argusflow.ai/api/agents/argusflow-translator/run \
  -H "Authorization: Bearer ak_live_..." \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "query": "Translate this to Spanish:\n\nThe project shipped on Tuesday."
    }
  }'

{
  "run_id": "run_a1b2c3...",
  "status": "completed",
  "output": "El proyecto se lanzó el martes.",
  "duration_ms": 8421,
  "cost_cents": 3,
  "tokens_used": 184,
  "version": "1.0.0"
}

From your codebase

// any-language pseudocode — Node example
const r = await fetch(
  "https://www.argusflow.ai/api/agents/argusflow-translator/run",
  {
    method: "POST",
    headers: {
      "Authorization": `Bearer ${process.env.ARGUSFLOW_API_KEY}`,
      "Content-Type": "application/json",
    },
    body: JSON.stringify({
      input: { query: "Translate this to Spanish:\n\n" + text },
    }),
  },
);
const { output } = await r.json();
console.log(output);

Auth modes

Three accepted credentials, in priority order:

• • Authorization: Bearer ak_live_... — API key from /dashboard/settings. Best for production.
• • Authorization: Bearer eyJ... — Supabase JWT. Best for in-browser calls from your own app.
• • Public agents (no require_auth) — no auth needed for try-out via the marketplace, rate-limited per IP.

4. Python SDK

For Python codebases, pip install cognimesh wraps the HTTP API with retries, streaming, and typed responses.

pip install cognimesh

from cognimesh import Client

client = Client(api_key="ak_live_...")

# One-shot
result = client.agents.run(
    "argusflow-translator",
    input={"query": "Translate this to Spanish:\n\nThe project shipped on Tuesday."},
)
print(result.output)

# Quote first
quote = client.agents.quote(
    "argusflow-translator",
    task_description="Translate a 3-paragraph email to Spanish",
)
print(f"Will cost {quote.price_cents} cents")

5. End-to-end: same agent from three surfaces

The same argusflow-translator agent, called three different ways. All three hit the same auction → quote → run pipeline, all three debit the same wallet, all three show up in your history with full chat transcripts.

# 1. Curl from a CI step
curl -X POST .../argusflow-translator/run -d '{"input":{"query":"..."}}'

# 2. Claude Desktop (the LLM picks the agent for you)
"Translate this email to Spanish"  →  argusflow_run_outcome auto-routes

# 3. Python script
client.agents.run("argusflow-translator", input={"query": "..."})

What's next

Today: MCP, HTTP API, Python SDK, JS via the public REST. Tomorrow:

• Slack — slash command + thread replies. Best fit for B2B teams that already live there.
• Telegram, Discord — bot tokens; same memory, just a different surface.
• Email + SMS / WhatsApp — async-friendly; reply to a prior thread to continue the conversation.
• Browser extension — right-click any selection → run with an agent.

Until those land, the API + MCP combination already covers any destination you can write a webhook for — Slack apps, Discord bots, Zapier zaps, n8n workflows, your own Lambda. The vision isn't blocked on us shipping every adapter.