AI agents: Skill, llms.txt and MCP

• For coding agents The Jusa Agent Skill Claude Code, Claude and any agent that reads Markdown learn the API, the money rules and your language's SDK.
• For any model llms.txt and llms-full.txt The docs as plain Markdown at the root of the site, to read or paste whole into a model's context.
• For assistants that act The MCP server Send airtime, buy ZESA and reward people from Claude, Cursor or VS Code, with a person approving above a threshold.

The Jusa Agent Skill

An Agent Skill is a folder an agent loads when a task needs it. SKILL.md says when to use it and how Jusa works: test and live keys, Jusa Credit at face value, an Idempotency-Key on every payment, how a send settles, ZESA meters, webhooks, typed errors. The files beside it are read only when a task needs them. The reference files are written from the code that decides them, so they change when the API does.

jusa/
├── SKILL.md
├── examples/
│   ├── dart.md
│   ├── dotnet.md
│   ├── go.md
│   ├── java.md
│   ├── node.md
│   ├── php.md
│   ├── python.md
│   └── ruby.md
└── reference/
    ├── api.md
    ├── errors.md
    ├── mcp.md
    ├── objects.md
    ├── recipes.md
    ├── test-mode.md
    └── webhooks.md

examples/ has the same job in each language, written as you would ship it: reward a completed survey once, then verify, dedupe and record the webhook that says how it went. Python, Node, PHP, Go, Java, C#, Dart, Ruby.

In Claude Code

Unzip it into your skills folder. Claude Code picks it up when a task mentions Jusa, airtime, ZESA or rewards; say use the jusa skill to be sure.

curl -fsSL https://jusa.localhost.co.zw/developers/ai/jusa-skill.zip -o /tmp/jusa-skill.zip
mkdir -p ~/.claude/skills && unzip -o /tmp/jusa-skill.zip -d ~/.claude/skills/

curl -fsSL https://jusa.localhost.co.zw/developers/ai/jusa-skill.zip -o /tmp/jusa-skill.zip
mkdir -p .claude/skills && unzip -o /tmp/jusa-skill.zip -d .claude/skills/
git add .claude/skills/jusa && git commit -m "Add the Jusa Agent Skill"

In Claude and Claude Desktop

1. Download jusa-skill.zip.
2. Open Settings, then Capabilities, and make sure code execution is on.
3. Under Skills, choose Upload skill and pick the zip. It holds one folder, jusa/, with SKILL.md at its top, as the upload expects.

In any other agent

Put SKILL.md where the agent reads its instructions (an AGENTS.md, a rules folder, a system prompt) and the reference/ and examples/ folders beside it, where it can open them. Its links are relative, so keep the folder as it is. Or give the agent /llms-full.txt.

llms.txt for any model

Two plain-text files at the root of this site, in the llms.txt format, written from the same pages and OpenAPI document as these docs.

curl -fsSL https://jusa.localhost.co.zw/llms-full.txt -o docs/jusa.md

The Jusa MCP server

For an assistant that acts for a person: send airtime, buy a ZESA token, reward someone, check the balance. The tools call the same code as the API, and an agent key holds them to limits a person sets.

Endpoint

POST https://jusa.localhost.co.zw/dev/v1/mcp

Transport

Streamable HTTP, answered as JSON. Stateless: no session, and GET is 405.

Key

Authorization: Bearer jusa_ak_…, agent keys only. OAuth is not offered yet.

Protocol

2025-06-18, 2025-03-26, 2024-11-05

First, an agent key

In the dashboard, switch to Test, open Keys and create an Agent key. Give it only the scopes its tools need, an approval threshold and, if you like, spend caps and the numbers it may reach without asking. With a secret key it is one call:

curl https://jusa.localhost.co.zw/dev/v1/keys \
  -H "Authorization: Bearer $JUSA_SECRET_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "kind": "ak",
    "label": "Support assistant",
    "scopes": ["catalog:read", "lookups:read", "sends:read", "sends:write", "credit:read"],
    "approval_threshold": {"USD": "5.00"},
    "spend_caps": {"USD": {"per_send": "5.00", "daily": "25.00", "monthly": "200.00"}},
    "recipient_allowlist": ["0772555001"]
  }'

Then, your client

export JUSA_AGENT_KEY=jusa_ak_test_…   # an agent key; test mode while you try it
claude mcp add --transport http --scope user jusa https://jusa.localhost.co.zw/dev/v1/mcp \
  --header "Authorization: Bearer $JUSA_AGENT_KEY"
claude mcp get jusa   # then /mcp inside Claude Code lists its tools

{
  "mcpServers": {
    "jusa": {
      "type": "http",
      "url": "https://jusa.localhost.co.zw/dev/v1/mcp",
      "headers": {
        "Authorization": "Bearer ${JUSA_AGENT_KEY}"
      }
    }
  }
}

{
  "permissions": {
    "allow": [
      "mcp__jusa__confirm_meter",
      "mcp__jusa__get_balance",
      "mcp__jusa__get_send",
      "mcp__jusa__list_products",
      "mcp__jusa__lookup_phone",
      "mcp__jusa__quote_zesa",
      "mcp__jusa__list_sends",
      "mcp__jusa__list_programs",
      "mcp__jusa__list_recipients",
      "mcp__jusa__get_status",
      "mcp__jusa__zesa_status",
      "mcp__jusa__search_docs"
    ],
    "ask": [
      "mcp__jusa__send_airtime",
      "mcp__jusa__buy_zesa",
      "mcp__jusa__create_reward",
      "mcp__jusa__send_data",
      "mcp__jusa__cancel_send",
      "mcp__jusa__resend_token"
    ]
  }
}

{
  "mcpServers": {
    "jusa": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://jusa.localhost.co.zw/dev/v1/mcp",
        "--header",
        "Authorization:${JUSA_AUTHORIZATION}"
      ],
      "env": {
        "JUSA_AUTHORIZATION": "Bearer jusa_ak_test_…"
      }
    }
  }
}

{
  "mcpServers": {
    "jusa": {
      "url": "https://jusa.localhost.co.zw/dev/v1/mcp",
      "headers": {
        "Authorization": "Bearer ${env:JUSA_AGENT_KEY}"
      }
    }
  }
}

{
  "inputs": [
    {
      "type": "promptString",
      "id": "jusa-agent-key",
      "description": "Jusa agent key (jusa_ak_test_… while you try it)",
      "password": true
    }
  ],
  "servers": {
    "jusa": {
      "type": "http",
      "url": "https://jusa.localhost.co.zw/dev/v1/mcp",
      "headers": {
        "Authorization": "Bearer ${input:jusa-agent-key}"
      }
    }
  }
}

curl -s https://jusa.localhost.co.zw/dev/v1/mcp \
  -H "Authorization: Bearer $JUSA_AGENT_KEY" \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc": "2.0", "id": 1, "method": "tools/list"}'

What keeps an agent in bounds

Every rule is checked on the server, under the same lock as the balance, whatever the client asks for.

The tools

18 tools; a key sees the ones its scopes allow. send_data is offered in live mode once live data bundles ship (test mode has simulated ones).

The skill's reference/mcp.md has all of this for an agent to read, and the AI agents recipe shows the approval flow end to end.