kalshi-markets

• MCP Servers are the standard way to build reusable toolsets for your AI Agents. But they are not the only way.
• MCP Servers come with a massive cost - instant context loss.
• When you have a single, or a few MCP Servers, this is not a big deal. But as you scale to many agents, many tools, and many contexts - this cost quickly becomes a bottleneck.
• So what are the alternatives that big players are using to build powerful, reusable, context preserving toolsets for their AI Agents?

_Here we explore 4 concrete approaches in this repo, all implementing access to Kalshi prediction market data._


`apps/1_mcp_server/` - MCP Server


`apps/2_cli/` - CLI


`apps/3_file_system_scripts/` - File System Scripts


`apps/4_skill/` - Skill


1. MCP Server

```bash

cp .mcp.testing .mcp.json

claude --mcp-config .mcp.json

prompt: "kalshi: get exchange status"

```

2. CLI

```bash

# or by agent

claude

prompt: "/prime_kalshi_cli_tools"

prompt: "kalshi: Get exchange status"

prompt: "kalshi: List events"

prompt: "kalshi: List events in JSON"

prompt: "kalshi: List events in JSON, limit 100"

# or by hand

cd apps/2_cli

uv sync

uv run kalshi status

uv run kalshi events

uv run kalshi events --json

uv run kalshi events --json --limit 100

```

3. File System Scripts

```bash

# by agent

claude

prompt: "/prime_file_system_scripts"

prompt: "kalshi: Get exchange status"

prompt: "kalshi: List events"

...

# or by hand

cd apps/3_file_system_scripts/scripts

uv run status.py

uv run *.py

```

4. Skill

```bash

cd apps/4_skill/

claude

prompt: "kalshi markets: Get exchange status"

prompt: "kalshi markets: search for events about 'best ai'" # Note this will trigger the cache build on first run which will take several minutes

...

```

• apps/1_mcp_server/ - MCP Server
• apps/2_cli/ - CLI
• apps/3_file_system_scripts/ - File System Scripts
• apps/4_skill/ - Skill

1. MCP Server (`apps/1_mcp_server/`)

Classic Model Context Protocol implementation

• ✅ Standardized integration - Works with any MCP-compatible client
• ✅ Tool discovery - Auto-exposes 15 tools to LLMs
• ✅ Clean abstractions - MCP protocol handles complexity
• ❌ Instant context loss - Every tool call loses conversational context
• ❌ Wrapper overhead - Delegates to CLI via subprocess

Architecture:

```

Claude/LLM → MCP Protocol → MCP Server → subprocess → CLI → Kalshi API

```

Key files:

• server.py - FastMCP server with 15 tool definitions
• Wraps CLI commands in MCP tool interface
• Each tool call is stateless

When to use: Building tools for multiple LLM clients, need standardized protocol, context loss is acceptable.

---

2. CLI (`apps/2_cli/`)

Direct HTTP API access via command-line interface

• ✅ Single source of truth - Direct API calls, no wrappers
• ✅ Dual output modes - Human-readable or pure JSON
• ✅ Smart caching - Pandas-based search with 6-hour TTL
• ✅ Minimal overhead - Direct httpx calls, no SDK
• ✅ Improved Context - Agent reads ~half as much context as the MCP Server

Architecture:

```

Claude → subprocess → CLI (13 commands) → Direct HTTP → Kalshi API

```

Key files:

• kalshi_cli/cli.py - All 13 commands (552 lines)
• kalshi_cli/modules/client.py - HTTP