MCP & Tools

Trading Agents interact with exchanges and local operations through MCP (Model Context Protocol) servers. MCP provides a standardized way for LLMs to call tools via JSON-RPC 2.0.

Architecture

Condor runs two MCP servers:

Server	Purpose
hummingbot-mcp	Trading operations via Hummingbot API (13 tools)
condor-mcp	Local operations: skills, domain agents, memory, notifications, routines, journals

Hummingbot MCP Tools

Account Management

Tool	Description
`setup_connector()`	Add/remove exchange API credentials
`configure_server()`	Switch between Hummingbot API servers

Portfolio & Holdings

Tool	Description
`get_portfolio_overview()`	Unified view: balances, positions, LP, orders

Trading Operations

Tool	Description
`set_account_position_mode_and_leverage()`	Configure perpetual trading settings
`search_history()`	Query historical orders, positions, LP
`get_market_data()`	Prices, candles, funding rates, order books

Executor Management

The primary trading interface for agents:

Tool	Description
`manage_executors()`	Create, search, stop trading executors

Supported executor types:

order_executor - Single limit/market orders
position_executor - Directional trades with triple barrier
grid_executor - Multi-level grid trading
dca_executor - Dollar-cost averaging
lp_executor - Liquidity provision

Actions: create, search, stop, get_logs

Bot Management

For advanced multi-strategy deployments:

Tool	Description
`manage_bots()`	Deploy and control controller-based bots
`manage_controllers()`	CRUD operations on controller templates

Market Data & Discovery

Tool	Description
`explore_dex_pools()`	Discover DEX/CLMM pools with filtering
`explore_geckoterminal()`	Free market data: networks, pools, OHLCV

Backtesting

Tool	Description
`run_backtest()`	Backtest a V2 controller config over historical data
`manage_backtest_tasks()`	Track and manage backtest tasks

DEX liquidity positions are created through manage_executors() (the lp_executor type), and Gateway itself is managed from the Condor dashboard (Settings → Gateway) rather than through a dedicated MCP tool.

Condor MCP Tools

Skills and domain agents

The routing model, and the part most worth knowing: before reaching for a raw tool, an assistant looks for a skill (a playbook, sometimes bound to a runnable routine) and then for a domain agent to hand the work to.

Tool	Description
`manage_skill()`	Read, search, list, create, edit and delete skills. Skills belong to the assistant, not to a user — they are the shared playbooks everyone using it gets
`consult()`	Ask a domain agent a question and block for its answer. The agent runs in the main process, where the runtime and server credentials live, and may pause on a user confirmation before acting
`delegate()`	The unattended sibling of `consult()`. Hands a goal-oriented task to a detached agent that runs until done and then notifies the user; returns a `task_id` to poll or stop

Authoring or debugging a routine goes through delegate() to a background Condor worker rather than being written by hand — the worker reads the routine cookbook, writes the routine, tests it, and reports back. Running an existing routine is not authoring: that is just manage_routines().

Memory and code

Tool	Description
`manage_memory()`	What the assistant remembers about you. Per-user, keyed to the calling user, with the writing agent recorded for audit
`run_code()`	Ad-hoc Python in the bot process, with the same primitives a routine gets — `context`, `client`, pandas, report builders, every `condor.*` module. The scratchpad below a routine: a snippet that failed is stored whole, read back, corrected and re-run

run_code() executes arbitrary Python inside the bot process. Granting it to a non-admin is the code_run permission in Settings → Admin, which spells out what it reaches before you grant it.

Notifications and routines

Tool	Description
`send_notification()`	Send Telegram messages (Markdown/HTML)
`manage_routines()`	Discover, run, create and edit routines

Trading agent operations

Tool	Description
`manage_trading_agent()`	Strategy CRUD and agent lifecycle (start/stop/pause)
`trading_agent_journal_read()`	Read the journal: recent entries, learnings, state
`trading_agent_journal_write()`	Write to the journal: actions, learnings, state

Environment

Tool	Description
`manage_servers()`	List accessible API servers
`get_user_context()`	Current user, active server, permissions
`get_available_models()`	Which models and providers are actually usable right now — installed CLI bridges, cloud keys present, reachable local servers and what they have loaded. Read-only, and never returns a key value, only whether its variable is set
`manage_notes()`	Key-value persistent storage

Progressive Disclosure

Many tools support step-by-step discovery:

Call with no parameters → See available options
Call with partial parameters → See next steps
Call with all parameters → Execute action

Example: Creating an executor

Agent: manage_executors()
Tool: Available executor types: order_executor, position_executor, grid_executor...

Agent: manage_executors(action="create", executor_type="grid_executor")
Tool: Required fields for grid_executor:
  - connector_name (string)
  - trading_pair (string)
  - ...

Agent: manage_executors(action="create", executor_type="grid_executor", config={...})
Tool: Grid executor created (ID: exec_123)

Configuration

Server Settings

MCP servers read configuration from environment variables or ~/.hummingbot_mcp/server.yml:

Variable	Default	Description
`HUMMINGBOT_API_URL`	`http://localhost:8000`	API server URL
`HUMMINGBOT_USERNAME`	`admin`	API username
`HUMMINGBOT_PASSWORD`	`admin`	API password
`HUMMINGBOT_TIMEOUT`	`30.0`	Request timeout (seconds)
`HUMMINGBOT_MAX_RETRIES`	`3`	Retry attempts

Runtime Configuration

Agents can switch servers at runtime:

Agent: configure_server()
Tool: Current server: localhost:8000 (admin)

Agent: configure_server(url="http://prod-server:8000", username="trader", password="xxx")
Tool: Connected to prod-server:8000

Error Handling

MCP servers provide contextual error messages:

Connection errors: Suggests checking server URL and Docker networking
Auth errors: Prompts to verify credentials via configure_server()
Validation errors: Shows required fields and valid values

All tools use retry logic (3 attempts with 2-second delays) for transient failures.

Running MCP Servers

MCP servers are started automatically when agents launch via ACP. For manual testing:

# Hummingbot MCP
python -m mcp_servers.hummingbot_api

# Condor MCP
python -m mcp_servers.condor

Adding Custom Tools

To add tools to an MCP server, create a function with the @mcp.tool() decorator:

# In mcp_servers/hummingbot_api/tools/custom.py

@mcp.tool()
async def my_custom_tool(param1: str, param2: int = 10) -> str:
    """
    Description of what this tool does.

    Args:
        param1: Description of param1
        param2: Description of param2 (default: 10)
    """
    # Implementation
    result = await do_something(param1, param2)
    return format_result(result)

from .tools.custom import my_custom_tool

Condor

Getting Started

Web UI Pages

The Condor Agent

Agent Architecture

Architecture

Hummingbot MCP Tools

Account Management

Portfolio & Holdings

Trading Operations

Executor Management

Bot Management

Market Data & Discovery

Backtesting

Condor MCP Tools

Skills and domain agents

Memory and code

Notifications and routines

Trading agent operations

Environment

Progressive Disclosure

Configuration

Server Settings

Runtime Configuration

Error Handling

Running MCP Servers

Adding Custom Tools

​Architecture

​Hummingbot MCP Tools

​Account Management

​Portfolio & Holdings

​Trading Operations

​Executor Management

​Bot Management

​Market Data & Discovery

​Backtesting

​Condor MCP Tools

​Skills and domain agents

​Memory and code

​Notifications and routines

​Trading agent operations

​Environment

​Progressive Disclosure

​Configuration

​Server Settings

​Runtime Configuration

​Error Handling

​Running MCP Servers

​Adding Custom Tools

Architecture

Hummingbot MCP Tools

Account Management

Portfolio & Holdings

Trading Operations

Executor Management

Bot Management

Market Data & Discovery

Backtesting

Condor MCP Tools

Skills and domain agents

Memory and code

Notifications and routines

Trading agent operations

Environment

Progressive Disclosure

Configuration

Server Settings

Runtime Configuration

Error Handling

Running MCP Servers

Adding Custom Tools