AI Coding Agents

Overview

LocalRouter can orchestrate AI coding agents (Claude Code, Gemini CLI, Codex, and more) as MCP tools through the Unified MCP Gateway. Any MCP client can spawn, interact with, and manage coding agent sessions — turning LocalRouter into a multi-agent orchestration hub.

Each client can be assigned one coding agent type. The selected agent is exposed through a unified set of MCP tools (default: AgentStart, AgentSay, AgentStatus, AgentList) that manage the full session lifecycle. The tool prefix is configurable.

Agent process management is powered by BloopAI/vibe-kanban's executors crate, providing robust process lifecycle management including kill_on_drop, graduated signal escalation, and for Claude Code: the full SDK control protocol with session resumption via --resume.

Coding agents are implicitly enabled when their binary is found on the system PATH. No manual configuration is required — install a supported agent and it's automatically available.

Sessions are strictly tied to the creating client. No cross-client session visibility or sharing.

Supported Agents

LocalRouter supports the following coding agents:

Agent	Binary	Description
Claude Code	`claude`	Anthropic's agentic coding tool
Gemini CLI	`gemini`	Google's AI coding assistant
Codex	`codex`	OpenAI's autonomous coding agent
Amp	`amp`	Sourcegraph's AI coding agent
Aider	`aider`	AI pair programming in terminal
Opencode	`opencode`	Open-source terminal AI assistant
Cursor	`cursor`	Cursor's CLI agent
Qwen Code	`qwen-code`	Alibaba's Qwen-based agent
GitHub Copilot	`gh`	GitHub Copilot's CLI extension
Droid	`droid`	Autonomous coding agent

An agent appears in the Coding Agents view as "installed" when its binary is found on the system PATH.

MCP Tool Interface

The selected coding agent is exposed through 4 unified MCP tools. The tool prefix is configurable (default: Agent). When the prefix ends with an alphanumeric character, suffixes are PascalCase (AgentStart). When it ends with a non-alphanumeric character like _, suffixes stay lowercase (agent_start).

Starting Sessions

Start (default: AgentStart) — Start a new coding session with an initial prompt.

Parameters:

prompt (required) — The initial task or prompt
workingDirectory — Working directory for the session
model — Model override (optional, agent default applies)
permissionMode — auto, supervised, or plan (default: agent's configured mode)

Returns a sessionId and status: "active". The session runs asynchronously — the tool returns immediately.

This is the only tool that triggers a gateway permission check (Allow/Ask/Off).

Sending Messages

Say (default: AgentSay) — Send a message to an existing session. Can also interrupt and/or resume.

Parameters:

sessionId (required) — The session ID
message — The message to send. If session is done/error, resumes with context preserved via --resume.
interrupt — If true, interrupts current work before sending the message. If true with no message, just stops the agent.
permissionMode — Switch permission mode

Behavior:

Message only, session active: sends the message to the agent (queued)
Message only, session ended: auto-resumes the session with context preserved
Interrupt only: gracefully interrupts the agent
Interrupt + message: interrupts, then resumes with the new message
Neither message nor interrupt: returns an error

Checking Status

Status (default: AgentStatus) — Check session status and retrieve recent output.

Parameters:

sessionId (required) — The session ID
outputLines — Recent output lines to return (default: 50)
wait — If true, blocks until the session needs attention (done, awaiting_input, error, interrupted) instead of returning immediately
timeoutSeconds — Max seconds to wait when wait=true (default: 300, max: 600)

Returns status (active, awaiting_input, done, error, interrupted), recent output lines, cost estimate, and turn count.

Use wait: true to block until the agent needs attention, instead of polling in a loop.

Responding to Questions

Interrupting Sessions

Listing Sessions

List (default: AgentList) — List all sessions for this client.

Parameters:

limit — Max sessions to return (default: 50)

Returns a list of sessions with their ID, working directory, display text, timestamp, and status.

Approval & Elicitation Flow

When a coding agent needs approval (tool use, plan review) or wants to ask a clarification question, LocalRouter routes it through one of three configurable approval modes:

Elicitation (Direct)

Polling (Fallback)

Session Lifecycle

Sessions follow a state machine:

Active — Agent is processing. Transitions to done on completion, error on failure, or interrupted on cancel.
Done — Agent finished successfully. Can be resumed via the say tool (context preserved with --resume).
Error — Agent encountered an error. Can be resumed via the say tool.
Interrupted — Agent was interrupted. Can be resumed via the say tool.

Any terminal state (done, error, interrupted) transitions back to active when a message is sent — the session auto-resumes with context preserved (for agents that support it, like Claude Code).

Per-Client Permissions

Coding agent access is controlled per-client:

Permission state — allow, ask, or off (configured in the client's Coding Agents tab)
Agent type — Which coding agent this client uses (one per client)

Permission behavior:

Allow — Client can start sessions freely, no gateway prompt
Ask — Gateway prompts the user on session creation (start tool), then free interaction
Off — Client has no access to coding agents

Skills

MCP via LLM