Skip to content

raine/consult-llm-mcp

BranchesTags

Repository files navigation

consult-llm-mcp

An MCP server that lets Claude Code consult stronger AI models (GPT-5.4, Gemini 3.1 Pro, DeepSeek Reasoner, MiniMax M2.7) when Sonnet has you running in circles and you need to bring in the heavy artillery. Supports multi-turn conversations.

> Still getting this error after your fix. Ask gemini
  E5108: Error executing lua: attempt to call method 'child' (a nil value)

⏺ Let me consult Gemini to understand the Neovim treesitter API changes:

⏺ consult-llm:consult_llm (MCP)(prompt: "Neovim plugin broken due to treesitter
                                API changes...", files: ["lua/testonly.lua"],
                                model: "gemini-2.5-pro")
  ⎿ The issue is that iter_matches() can return nil nodes in newer Neovim
    versions. Switch to iter_captures() which never returns nil nodes, and use
    named_child(0) instead of child(0) to get only meaningful syntax nodes.
    … +93 lines (ctrl+r to expand)

⏺ The LLM identified the issue - iter_matches() API changed. Switching to
  iter_captures() approach now.

Quick start · Configuration · Skills · Monitor TUI · Why MCP? · Changelog

Features

  • Query powerful AI models (GPT-5.4, Gemini 3.1 Pro, DeepSeek Reasoner, MiniMax M2.7) with relevant files as context
  • Include git changes for code review
  • Comprehensive logging with cost estimation (if using API)
  • Monitor TUI: Real-time dashboard for watching active consultations
  • Gemini CLI backend: Use the gemini CLI for Gemini models
  • Codex CLI backend: Use the codex CLI for OpenAI models
  • Cursor CLI backend: Use the cursor-agent CLI to route GPT and Gemini models through a single tool
  • OpenCode CLI backend: Use opencode CLI with Copilot, OpenRouter, or any of 75+ providers
  • Multi-turn conversations: Resume CLI sessions across requests with thread_id
  • Web mode: Copy formatted prompts to clipboard for browser-based LLM services
  • Skills: Multi-LLM debate, collaboration, and consultation workflows
  • Less is more: Single MCP tool to not clutter the context

consult-llm-monitor screenshot

Quick start

  1. Add to Claude Code (choose one):

    With npx (no install required):

    claude mcp add consult-llm \
  -e CONSULT_LLM_GEMINI_BACKEND=gemini-cli \
  -e CONSULT_LLM_OPENAI_BACKEND=codex-cli \
  -- npx -y consult-llm-mcp

    This is the recommended setup. Uses Gemini CLI and Codex CLI. No API keys required, just gemini login and codex login.

    With binary (comes with the monitor TUI, no Node.js required):

    curl -fsSL https://raw.githubusercontent.com/raine/consult-llm-mcp/main/scripts/install.sh | bash
    claude mcp add consult-llm \
  -e CONSULT_LLM_GEMINI_BACKEND=gemini-cli \
  -e CONSULT_LLM_OPENAI_BACKEND=codex-cli \
  -- consult-llm-mcp

    For global availability across projects, add --scope user.

    Using API keys instead of CLI backends:

    claude mcp add consult-llm \
  -e OPENAI_API_KEY=your_openai_key \
  -e GEMINI_API_KEY=your_gemini_key \
  -e DEEPSEEK_API_KEY=your_deepseek_key \
  -e MINIMAX_API_KEY=your_minimax_key \
  -- npx -y consult-llm-mcp

  2. Verify connection with /mcp:

    ❯ 1. consult-llm            ✔ connected

  3. Ask a question:

    "Consult Gemini about how to fix the race condition in server.ts"

Example workflows

Some real-world examples. Click to expand.

Explain the problem, and tell CC to consult a smarter LLM 
  <Some long explanation of the issue here>

  Consult LLM for assistance, use gemini

⏺ I see the issue! The problem is that the frontend is trying to use the hint
  for "isn't" but the server generated a hint for "isn't it" (the full phrase).
  This is a mismatch between how the frontend segments text and how the server
  generates hints.

  Let me consult the LLM to get assistance on this segmentation mismatch issue.

⏺ Write(hint-segmentation-issue.md)
  ⎿  Wrote 53 lines to hint-segmentation-issue.md
     # Hint Segmentation Mismatch Issue
     ## Problem Description
     There's a mismatch between how the frontend segments text for clickable
     words and how the backend generates hints. This causes hint unlocking to
     fail when the segments don't match exactly.
     ## Example Case
     **Original Text**: "Hello! Nice weather, isn't it?"
     … +43 lines (ctrl+r to expand)

⏺ consult-llm:consult_llm (MCP)(model: "gemini-2.5-pro", files: ["/project/hint
                                -segmentation-issue.md","/project/web/src/compo
                                nents/HintsText.tsx","/project/web/src/utils/hi
                                nts.ts","/project/server/src/routes/public/hint
                                s.ts","/project/server/src/llm/hints.ts"])
  ⎿ An excellent analysis of the problem. The root cause has been correctly
    identified: a race condition within the `handleConfirmGenerateHints`
    function in `HintsText.tsx`. The component state updates in the background
    after `await generateHints()`, but the function continues executing with a
    stale `activeSegment` variable, causing it to request a hint for a smaller
    text segment ( e.g., "isn't") instead of the full, newly-available hint
    phrase (e.g., "isn't it").
    … +607 lines (ctrl+r to expand)

⏺ Perfect! The LLM identified the core issue - there's a race condition in
  handleConfirmGenerateHints where the function uses stale state after async
  operations. The solution is to re-target the hint key after generation to
  find the correct matching hint.

  Let me implement this solution:

⏺ Update Todos
  ⎿  ☐ Fix hint segmentation mismatch by implementing smarter hint key
       targeting in handleConfirmGenerateHints
Using web mode to copy prompts to clipboard for browser-based LLMs 
> I need help implementing a -b/--background flag for my Rust CLI. Use consult LLM with web mode

⏺ I'll use web mode to copy the formatted prompt to your clipboard so you can
  paste it into any browser-based LLM service.

⏺ consult-llm:consult_llm (MCP)(prompt: "What is the best approach to implement a
                                 -b/--background flag for the \"workmux add\" command that
                                 creates a tmux window in the background without switching to
                                 it?\n\nThe current implementation automatically switches to
                                 the newly created window via `tmux::select_window()` in the
                                 `setup_environment()` function. I need to make this behavior
                                 conditional based on a flag passed from the
                                 CLI.\n\nConsider:\n1. Where should the background flag be
                                 threaded through the call chain?\n2. Should I modify the
                                 existing SetupOptions struct or add a parameter?\n3. What's
                                 the most maintainable approach given the codebase
                                 structure?\n4. Are there any edge cases to consider (e.g.,
                                 should select_pane also be conditional)?", files:
                                 ["src/cli.rs","src/workflow.rs","src/tmux.rs"], web_mode:
                                 true)
  ⎿  ✓ Prompt copied to clipboard!

     Please paste it into your browser-based LLM service and share the response
     here before I proceed with any implementation.

     Note: File paths were included:
       - src/cli.rs
       - src/workflow.rs
       - src/tmux.rs

> [User pastes the response from the external LLM]

⏺ Thank you! Based on the recommendations, I'll now implement the solution...

This is useful when:

  • You want to use a free browser-based LLM service instead of API credits
  • You prefer a specific LLM's web interface
  • You want to review the full prompt before submitting it
Debugging Neovim plugin treesitter API breakage 
⏺ Fixed! Added nil checks for both node and child and updated the iter_matches call to handle the newer API properly.

> Still not working:

  E5108: Error executing lua /Users/raine/code/testonly.nvim/lua/testonly.lua:92: attempt to call method 'child' (a nil value)
  stack traceback:
          /Users/raine/code/testonly.nvim/lua/testonly.lua:92: in function 'reset_all_exclusive'
          /Users/raine/code/testonly.nvim/lua/testonly.lua:110: in function 'reset_all_exclusive'
          [string ":lua"]:1: in main chunk

  Ask gemini

⏺ Let me consult with a more powerful AI to understand the current Neovim treesitter API changes:

⏺ consult-llm:consult_llm (MCP)(prompt: "I'm debugging a Neovim plugin that's broken due to treesitter API changes. The error is \"attempt to call method 'child' (a nil value)\" on line 92. The
                               code is trying to access child nodes from treesitter matches but the API seems to have changed in newer Neovim versions.\n\nHere's the problematic
                               code:\n```lua\nfor _, match, _ in parsed_query:iter_matches(root, bufnr, start_row, end_row) do\n    for id, node in pairs(match) do\n        local name =
                               parsed_query.captures[id]\n        if name == \"call\" and node then\n            local child = node:child(0)  -- ERROR HERE\n            if child then\n
                                    local text = vim.treesitter.get_node_text(child, bufnr)\n                -- ... rest of logic\n            end\n        end\n    end\nend\n```\n\nWhat's the
                               correct way to iterate over treesitter query matches and access child nodes in current Neovim versions? What API changes should I be aware of?", files:
                               ["/Users/raine/code/testonly.nvim/lua/testonly.lua"])
  ⎿  Problem
     -------

     … +93 lines (ctrl+r to expand)

⏺ Perfect! The issue is that iter_matches() can return nil nodes in the match table. Let me fix this by switching to the recommended iter_captures() approach:

⏺ Update(lua/testonly.lua)
  ⎿  Updated lua/testonly.lua with 7 additions and 10 removals
       85      local root = parser:parse()[1]:root()
       86      local start_row, _, end_row, _ = root:range()
       87
       88 -    for _, match, _ in parsed_query:iter_matches(root, bufnr, start_row, end_row) do
       89 -      for id, node in pairs(match) do
       90 -        local name = parsed_query.captures[id]
       91 -        if name == "call" and node then
       92 -          local child = node:child(0)
       93 -          if child then
       94 -            local text = vim.treesitter.get_node_text(child, bufnr)
       88 +    for id, node, _ in parsed_query:iter_captures(root, bufnr, start_row, end_row) do
       89 +      if parsed_query.captures[id] == "call" then
       90 +        local first_child = node:named_child(0)
       91 +        if first_child then
       92 +          local text = vim.treesitter.get_node_text(first_child, bufnr)
       93
       94 -            if text == "it.only" or text == "describe.only" then
       95 -              toggle_test_exclusive(child)
       96 -            end
       94 +          if text == "it.only" or text == "describe.only" then
       95 +            toggle_test_exclusive(first_child)
       96            end
       97          end
       98        end

⏺ Fixed! The key changes:

  1. Switched from iter_matches() to iter_captures() which never returns nil nodes
  2. Used named_child(0) instead of child(0) to get only meaningful syntax nodes (not punctuation)
  3. Simplified the iteration logic since iter_captures() directly gives us the node and capture ID

  Your <leader>tr mapping should now work properly with the current Neovim treesitter API.
Using slash command to consult multiple LLMs for comparison 
> /consult The branch completions should dynamic. Git should be called only when
using tab to autocomplete. Ask gemini and codex

⏺ I'll consult both Gemini and Codex about making the branch completions dynamic (only
  calling git when tab-completing). Let me gather the relevant context first.

⏺ consult-llm - consult_llm (MCP)(prompt: "I have a Rust CLI tool using clap that needs dynamic shell completions. Currently, the branch completions are generated at completion script
                                 generation time via `possible_values()`. Instead, I need the git operations to only run when the user presses TAB to autocomplete.\n\nCurrent setup:\n1.
                                 `WorktreeBranchParser` implements `TypedValueParser` with `possible_values()` that calls git\n2. Shell completions are generated via clap_complete's
                                 `generate()` function\n3. The completions are sourced at shell startup: `source <(workmux completions zsh)`\n\nThe problem: Running `workmux completions
                                 zsh` takes ~210ms because it runs git operations during completion generation, not during actual tab completion.\n\nHow can I implement truly dynamic
                                 completions where git is only called when the user presses TAB? What's the best approach with clap and clap_complete for zsh?", model:
                                 "gemini-3-pro-preview", files: ["/Users/raine/code/workmux/src/cli.rs","/Users/raine/code/workmux/src/git.rs"], web_mode: false)
  ⎿ The root cause of the slow startup (~210ms) is that `clap_complete::generate()` executes your `possible_values()` logic (git operations) *while generating the script*. This bake
    s the *current* branches into the static script, which is both slow (happens every time you `source` it) and incorrect (branches become stale).
    … +116 lines (ctrl+o to expand)


⏺ consult-llm - consult_llm (MCP)(prompt: "I have a Rust CLI tool using clap that needs dynamic shell completions. Currently, the branch completions are generated at completion script
                                 generation time via `possible_values()`. Instead, I need the git operations to only run when the user presses TAB to autocomplete.\n\nCurrent setup:\n1.
                                 `WorktreeBranchParser` implements `TypedValueParser` with `possible_values()` that calls git\n2. Shell completions are generated via clap_complete's
                                 `generate()` function\n3. The completions are sourced at shell startup: `source <(workmux completions zsh)`\n\nThe problem: Running `workmux completions
                                 zsh` takes ~210ms because it runs git operations during completion generation, not during actual tab completion.\n\nHow can I implement truly dynamic
                                 completions where git is only called when the user presses TAB? What's the best approach with clap and clap_complete for zsh?", model:
                                 "gpt-5.1-codex-max", files: ["/Users/raine/code/workmux/src/cli.rs","/Users/raine/code/workmux/src/git.rs"], web_mode: false)
  ⎿ **Findings**
    - `WorktreeBranchParser::possible_values` (`cli.rs:41-72`) runs git during script generation, so `workmux completions zsh` always pays the git cost (~210ms) even though the resu
    lt is only needed at tab time.
    … +77 lines (ctrl+o to expand)


⏺ Summary from Gemini and Codex

  Both agree on the same solution:

  Root Cause

  clap_complete::generate() calls possible_values() during script generation, not at tab-completion time. This means git runs every shell startup.

...

This example shows using the /consult slash command to ask multiple LLMs (Gemini and Codex) about the same problem in parallel and compare their responses. Both LLMs independently arrived at the same solution, providing confidence in the approach.

Backends

Each model is routed to a backend — either an API endpoint or a CLI tool.

Backend Description When to use
API Queries LLM APIs directly You have API keys and want the simplest setup
Gemini CLI Shells out to gemini CLI Free quota (Gemini), existing subscriptions, or prefer CLI tools
Codex CLI Shells out to codex CLI OpenAI models via Codex subscription
Cursor CLI Shells out to cursor-agent CLI Route GPT and Gemini through one tool
OpenCode CLI Shells out to opencode CLI Use Copilot subscription, OpenCode's 75+ providers
Web Copies prompt to clipboard You prefer browser UIs or want to review prompts

API (default)

The default backend. Requires API keys configured via environment variables. See Configuration for details.

CLI backends

Instead of making API calls, shell out to local CLI tools. The CLI tools can explore the codebase themselves, so you don't need to pass all relevant files as context, but it helps.

Gemini CLI

Use Gemini's local CLI to take advantage of Google's free quota or use your Google AI Pro subscription.

Requirements:

  1. Install the Gemini CLI
  2. Authenticate via gemini login

Setup:

claude mcp add consult-llm -e CONSULT_LLM_GEMINI_BACKEND=gemini-cli -- npx -y consult-llm-mcp

Codex CLI

Use OpenAI's Codex CLI for OpenAI models.

Requirements:

  1. Install the Codex CLI
  2. Authenticate via codex login

Setup:

claude mcp add consult-llm -e CONSULT_LLM_OPENAI_BACKEND=codex-cli -- npx -y consult-llm-mcp

Tip

Reasoning effort defaults to high. Override with -e CONSULT_LLM_CODEX_REASONING_EFFORT=xhigh. Options: none, minimal, low, medium, high, xhigh.

Cursor CLI

Use Cursor's agent CLI to route GPT and Gemini models through one tool.

Requirements:

  1. Install the Cursor agent CLI (cursor-agent in PATH)

Setup:

# Route GPT models through Cursor CLI
claude mcp add consult-llm -e CONSULT_LLM_OPENAI_BACKEND=cursor-cli -- npx -y consult-llm-mcp

# Route Gemini models through Cursor CLI
claude mcp add consult-llm -e CONSULT_LLM_GEMINI_BACKEND=cursor-cli -- npx -y consult-llm-mcp

# Route everything through Cursor CLI
claude mcp add consult-llm \
  -e CONSULT_LLM_OPENAI_BACKEND=cursor-cli \
  -e CONSULT_LLM_GEMINI_BACKEND=cursor-cli \
  -e CONSULT_LLM_ALLOWED_MODELS="gemini-3.1-pro-preview,gpt-5.3-codex" \
  -- npx -y consult-llm-mcp

Shell command permissions:

Cursor CLI runs with --mode ask, which blocks shell commands by default. If your prompts involve tools that need to run commands (e.g. git diff for code review), allow them in ~/.cursor/cli-config.json:

{
  "permissions": {
    "allow": ["Shell(git diff*)", "Shell(git log*)", "Shell(git show*)"],
    "deny": []
  }
}

Glob patterns are supported. The deny list takes precedence over allow.

OpenCode CLI

Use OpenCode as a backend to route models through any of its 75+ supported providers — including GitHub Copilot, OpenRouter, and local models via Ollama.

Requirements:

  1. Install OpenCode
  2. Configure providers via opencode providers

Setup:

# Route MiniMax models through OpenCode
claude mcp add consult-llm \
  -e CONSULT_LLM_MINIMAX_BACKEND=opencode \
  -- npx -y consult-llm-mcp

# Route OpenAI models through Copilot subscription
claude mcp add consult-llm \
  -e CONSULT_LLM_OPENAI_BACKEND=opencode \
  -e CONSULT_LLM_OPENCODE_OPENAI_PROVIDER=copilot \
  -- npx -y consult-llm-mcp

# Route everything through OpenCode
claude mcp add consult-llm \
  -e CONSULT_LLM_OPENAI_BACKEND=opencode \
  -e CONSULT_LLM_GEMINI_BACKEND=opencode \
  -e CONSULT_LLM_DEEPSEEK_BACKEND=opencode \
  -e CONSULT_LLM_MINIMAX_BACKEND=opencode \
  -- npx -y consult-llm-mcp

The executor maps model IDs to OpenCode's provider/model format automatically. For example, MiniMax-M2.7 becomes opencode run --model minimax/MiniMax-M2.7.

Provider prefix overrides:

By default, each provider family maps to its natural OpenCode provider ID (openai, google, deepseek, minimax). Override with per-family env vars when you want to route through a different OpenCode provider:

  • CONSULT_LLM_OPENCODE_OPENAI_PROVIDER — default: openai
  • CONSULT_LLM_OPENCODE_GEMINI_PROVIDER — default: google
  • CONSULT_LLM_OPENCODE_DEEPSEEK_PROVIDER — default: deepseek
  • CONSULT_LLM_OPENCODE_MINIMAX_PROVIDER — default: minimax
  • CONSULT_LLM_OPENCODE_PROVIDER — global fallback for all families

For example, CONSULT_LLM_OPENCODE_OPENAI_PROVIDER=copilot turns gpt-5.2 into opencode run --model copilot/gpt-5.2.

Multi-turn conversations

CLI backends support multi-turn conversations via the thread_id parameter. The first response includes a [thread_id:xxx] prefix. Pass that ID in follow-up requests to continue the conversation with full context from prior turns.

This works with all CLI backends (Gemini CLI, Codex CLI, Cursor CLI). Each uses its own session/thread mechanism, but all are passed through the same thread_id parameter.

> Ask codex what's the best caching strategy for our read-heavy API

⏺ consult-llm - consult_llm (MCP)(prompt: "What's the best caching strategy for a read-heavy API?",
                                   model: "gpt-5.3-codex")
  ⎿  [thread_id:thread_b1ff711...]

     Redis with TTL-based invalidation. Set short TTLs for volatile data ...

⏺ Codex recommends Redis with TTL-based invalidation — but our payment
  endpoints need strong consistency. Let me push back on that.

⏺ consult-llm - consult_llm (MCP)(prompt: "What if we need strong consistency guarantees?
                                            Does that change your recommendation?",
                                   model: "gpt-5.3-codex",
                                   thread_id: "thread_b1ff711...")
  ⎿  [thread_id:thread_b1ff711...]

     Yes — with strong consistency, the TTL approach breaks down. Consider
     write-through caching or cache invalidation on writes instead ...

⏺ Good — Codex revised its recommendation based on the consistency
  constraint. I'll use write-through caching for the payment endpoints.

See the debate skills for multi-LLM workflows that use thread_id to maintain context across debate rounds.

Web mode

Copies the formatted prompt to clipboard instead of querying an LLM. Paste into any browser-based LLM (ChatGPT, Claude.ai, Gemini, etc.).

When to use: Prefer a specific web UI, want to review the prompt first, or don't have API keys.

Workflow:

  1. Ask Claude to "use consult LLM with web mode"
  2. Paste into your browser-based LLM
  3. Paste the response back into Claude Code

See the "Using web mode..." example above for a concrete transcript.

Configuration

Environment variables

  • OPENAI_API_KEY - Your OpenAI API key (required for OpenAI models in API mode)
  • GEMINI_API_KEY - Your Google AI API key (required for Gemini models in API mode)
  • DEEPSEEK_API_KEY - Your DeepSeek API key (required for DeepSeek models)
  • MINIMAX_API_KEY - Your MiniMax API key (required for MiniMax models)
  • CONSULT_LLM_DEFAULT_MODEL - Override the default model (optional)
    • Accepts selectors (gemini, openai, deepseek, minimax) or exact model IDs (gpt-5.4, gemini-3.1-pro-preview, etc.)
    • Selectors are resolved to the best available model at startup
  • CONSULT_LLM_GEMINI_BACKEND - Backend for Gemini models (optional)
    • Options: api (default), gemini-cli, cursor-cli, opencode
  • CONSULT_LLM_OPENAI_BACKEND - Backend for OpenAI models (optional)
    • Options: api (default), codex-cli, cursor-cli, opencode
  • CONSULT_LLM_DEEPSEEK_BACKEND - Backend for DeepSeek models (optional)
    • Options: api (default), opencode
  • CONSULT_LLM_MINIMAX_BACKEND - Backend for MiniMax models (optional)
    • Options: api (default), opencode
  • CONSULT_LLM_ALLOWED_MODELS - Restrict which concrete models can be used (optional)
    • Comma-separated list, e.g., gpt-5.4,gemini-3.1-pro-preview
    • Selectors resolve against this list — e.g., if only gemini-2.5-pro is allowed, the gemini selector resolves to it
    • Useful when a backend doesn't support all models (e.g., Cursor CLI)
    • See Tips for usage examples
  • CONSULT_LLM_EXTRA_MODELS - Add models not in the built-in list (optional)
    • Comma-separated list, e.g., grok-3,kimi-k2.5
    • Merged with built-in models and included in the tool schema
    • Useful for newly released models with a known provider prefix (gpt-, gemini-, deepseek-, MiniMax-)
  • CONSULT_LLM_CODEX_REASONING_EFFORT - Configure reasoning effort for Codex CLI (optional, default: high)
    • See Codex CLI for details and available options
  • CONSULT_LLM_OPENCODE_PROVIDER - Global OpenCode provider prefix (optional)
    • Overrides the default provider ID for all families when using the opencode backend
    • See OpenCode CLI for details and per-family overrides
  • CONSULT_LLM_SYSTEM_PROMPT_PATH - Custom path to system prompt file (optional)
    • Overrides the default ~/.consult-llm-mcp/SYSTEM_PROMPT.md location
    • Useful for project-specific prompts
  • CONSULT_LLM_NO_UPDATE_CHECK - Disable automatic update checking on server startup (optional)
    • Set to 1 to disable
    • By default, the server checks for new versions in the background every 24 hours and logs a notice when an update is available
    • Only applies to binary installs — npm installs are never checked
  • MCP_DEBUG_STDIN - Log raw JSON-RPC messages received on stdin (optional)
    • Set to 1 to enable
    • Logs every message as RAW RECV entries and poll timing gaps as STDIN POLL entries in mcp.log
    • Useful for debugging transport-level issues

Custom system prompt

You can customize the system prompt used when consulting LLMs by creating a SYSTEM_PROMPT.md file in ~/.consult-llm-mcp/:

npx consult-llm-mcp init-prompt

This creates a placeholder file with the default system prompt that you can edit to customize how the consultant LLM behaves. The custom prompt is read on every request, so changes take effect immediately without restarting the server.

When a custom prompt file exists, it acts as a full override — task_mode overlays are not applied on top. To revert to the default prompt with task_mode support, simply delete the SYSTEM_PROMPT.md file.

Custom prompt path

Use CONSULT_LLM_SYSTEM_PROMPT_PATH to override the default prompt file location. This is useful for project-specific prompts that you can commit to your repository:

claude mcp add consult-llm \
  -e GEMINI_API_KEY=your_key \
  -e CONSULT_LLM_SYSTEM_PROMPT_PATH=/path/to/project/.consult-llm-mcp/SYSTEM_PROMPT.md \
  -- npx -y consult-llm-mcp

Tips

Controlling which models are used

The model parameter accepts selectors (gemini, openai, deepseek) that the server resolves to the best available concrete model. When no model is specified, the server uses CONSULT_LLM_DEFAULT_MODEL or its built-in fallback.

Selector resolution order (first available wins):

Selector Priority
gemini gemini-3.1-pro-preview → gemini-3-pro-preview → gemini-2.5-pro
openai gpt-5.4 → gpt-5.3-codex → gpt-5.2 → gpt-5.2-codex
deepseek deepseek-reasoner

Restricting models with CONSULT_LLM_ALLOWED_MODELS:

If your backend doesn't support all models (e.g., Cursor CLI can't use gpt-5.4), use CONSULT_LLM_ALLOWED_MODELS to filter. Selectors will automatically resolve to the best model within the allowed list:

# Only allow codex models through Cursor CLI
claude mcp add consult-llm \
  -e CONSULT_LLM_OPENAI_BACKEND=cursor-cli \
  -e CONSULT_LLM_ALLOWED_MODELS='gpt-5.3-codex,gemini-3.1-pro-preview' \
  -- npx -y consult-llm-mcp
# "openai" selector → gpt-5.3-codex (gpt-5.4 filtered out)

MCP tool: consult_llm

The server provides a single tool called consult_llm for asking powerful AI models complex questions.

Parameters

  • prompt (required): Your question or request for the consultant LLM

  • files (optional): Array of file paths to include as context

    • All files are added as context with file paths and code blocks

  • model (optional): Model selector or exact model ID

    • Selectors: gemini, openai, deepseek — the server resolves to the best available model for each family
    • Exact model IDs (gpt-5.4, gemini-3.1-pro-preview, etc.) are also accepted as an advanced override
    • When omitted, the server uses the configured default

  • task_mode (optional): Controls the system prompt persona. The calling LLM should choose based on the task:

    • general (default): Neutral base prompt that defers to the user prompt
    • review: Critical code reviewer — bugs, security, performance, anti-patterns
    • debug: Focused troubleshooter — root cause analysis, execution tracing, ignores style issues
    • plan: Constructive architect — trade-offs, alternatives, always includes a final recommendation
    • create: Generative writer — docs, content, polished output

  • web_mode (optional): Copy prompt to clipboard instead of querying LLM

    • Default: false
    • When true, the formatted prompt (including system prompt and file contents) is copied to clipboard for manual pasting into browser-based LLM services

  • thread_id (optional): Resume a multi-turn conversation

    • Works with CLI backends (Codex CLI, Gemini CLI, Cursor CLI)
    • The first response includes a [thread_id:xxx] prefix — pass that ID back as thread_id in follow-up requests to maintain conversation context

  • git_diff (optional): Include git diff output as context

    • files (required): Specific files to include in diff
    • repo_path (optional): Path to git repository (defaults to current directory)
    • base_ref (optional): Git reference to compare against (defaults to HEAD)

Supported models

  • gemini-2.5-pro: Google's Gemini 2.5 Pro
  • gemini-3-pro-preview: Google's Gemini 3 Pro Preview
  • gemini-3.1-pro-preview: Google's Gemini 3.1 Pro Preview
  • deepseek-reasoner: DeepSeek's reasoning model
  • MiniMax-M2.7: MiniMax's M2.7 reasoning model (204K context)
  • gpt-5.4: OpenAI's GPT-5.4 model
  • gpt-5.2: OpenAI's GPT-5.2 model
  • gpt-5.3-codex: OpenAI's Codex model based on GPT-5.3
  • gpt-5.2-codex: OpenAI's Codex model based on GPT-5.2

Logging

All prompts and responses are logged to $XDG_STATE_HOME/consult-llm-mcp/mcp.log (defaults to ~/.local/state/consult-llm-mcp/mcp.log) with:

  • Tool call parameters
  • Full prompts and responses
  • Token usage and cost estimates
Example 
[2025-06-22T20:16:04.673Z] TOOL CALL: consult_llm
Arguments: {
  "files": [
    "refactor-analysis.md",
    "src/main.ts",
    "src/schema.ts",
    "src/config.ts",
    "src/llm.ts",
    "src/llm-cost.ts"
  ],
  "model": "deepseek-reasoner"
}
================================================================================
[2025-06-22T20:16:04.675Z] PROMPT (model: deepseek-reasoner):
## Relevant Files

### File: src/main.ts

...

Please provide specific suggestions for refactoring with example code structure
where helpful.
================================================================================
[2025-06-22T20:19:20.632Z] RESPONSE (model: deepseek-reasoner):
Based on the analysis, here are the key refactoring suggestions to improve
separation of concerns and maintainability:

...

This refactoring maintains all existing functionality while significantly
improving maintainability and separation of concerns. The new structure makes
it easier to add features like new LLM providers, additional context sources,
or alternative prompt formats.

Tokens: 3440 input, 5880 output | Cost: $0.014769 (input: $0.001892, output: $0.012877)

Monitor

consult-llm-monitor is a real-time TUI dashboard for watching active consultations across all running MCP server instances. It shows what's being consulted, which models are in use, and how long each request takes.

consult-llm-monitor demo

The monitor binary is included when you install via the install script (same script that installs consult-llm-mcp).

consult-llm-monitor

The main table view shows two panels: active server instances with their in-flight consultations, and a history log of completed consultations with timestamps, models, durations, and token counts.

Press Enter on any consultation to open the detail view with the full event log - prompt, response with syntax-highlighted markdown, tool calls, and token usage. Press ? for keyboard shortcuts.

Activation methods

1. No custom activation (simplest)

When you add an MCP to Claude Code, the tool's schema is injected into the agent's context. This allows Claude to infer when to call the MCP from natural language (e.g., "ask gemini about..."). Works out of the box, but you have less control over how the MCP is invoked.

2. Skills

Automatically triggers when Claude detects matching intent. Like slash commands, supports custom instructions (e.g., always gathering relevant files), but not always reliably triggered. See the consult skill below.

Recommendation: Start with no custom activation. Use skills if you need custom instructions for how the MCP is invoked.

Skills

Installing skills

Install all skills globally with a single command:

curl -fsSL https://raw.githubusercontent.com/raine/consult-llm-mcp/main/scripts/install-skills | bash

This installs skills for all detected platforms:

  • Claude Code → ~/.claude/skills/
  • OpenCode → ~/.config/opencode/skills/
  • Codex → ~/.codex/skills/

To uninstall:

curl -fsSL https://raw.githubusercontent.com/raine/consult-llm-mcp/main/scripts/install-skills | bash -s uninstall

consult

An example Claude Code skill that uses the consult_llm MCP tool to create commands like "ask gemini" or "ask codex". See skills/consult/SKILL.md for the full content.

Type "ask gemini about X" or "ask codex about X" in Claude Code. This is not strictly necessary since Claude can infer from the schema that "ask gemini" should call this MCP, but it gives more precise control over how the agent calls this MCP.

collab

Collaborative ideation. Gemini and Codex independently brainstorm ideas, then build on each other's suggestions across multiple rounds. Unlike debate, the tone is cooperative — refining and combining rather than critiquing. Claude synthesizes the strongest ideas into a plan and implements. See skills/collab/SKILL.md.

> /collab how should we handle offline sync for the mobile app

collab-vs

Claude brainstorms with one LLM. Claude and an opponent (Gemini or Codex) take turns building on each other's ideas. Like collab, but Claude participates directly instead of moderating. See skills/collab-vs/SKILL.md.

> /collab-vs --gemini how should we handle offline sync for the mobile app

debate

Claude moderates, two LLMs debate. Gemini and Codex independently propose approaches, then critique each other's proposals. Claude synthesizes the best ideas and implements. See skills/debate/SKILL.md.

> /debate design the multi-tenant isolation strategy

debate-vs

Claude participates as a debater against one opponent LLM (Gemini or Codex) through multiple rounds. Claude forms its own position, then debates back and forth before synthesizing and implementing. See skills/debate-vs/SKILL.md.

> /debate-vs --gemini design the multi-tenant isolation strategy

Updating

Binary installs:

consult-llm-mcp update

Downloads the latest release from GitHub with SHA-256 checksum verification. If consult-llm-monitor is found alongside the binary, it's updated too.

The server also checks for updates in the background on startup (every 24 hours) and logs a notice when a newer version is available. Disable with CONSULT_LLM_NO_UPDATE_CHECK=1.

Why MCP and not CLI?

The server maps one model parameter onto five backends (OpenAI API, Gemini API, Gemini CLI, Codex CLI, Cursor CLI) with different commands, streaming formats, output schemas, file handling, and resume semantics. Doing this through agent Bash calls would push all of that per-provider plumbing into the agent or a wrapper script/CLI.

MCP also sidesteps shell escaping. Prompts contain code with backticks, $, and quotes. Passing one model's code-heavy response into another call breaks bash quoting and requires temp files. MCP passes structured JSON instead.

Multi-turn workflows add more friction as a CLI. To continue a conversation, the agent needs to find a session ID in the CLI's output and pass it back as a flag on the next invocation. With MCP, the agent passes thread_id as a parameter and the server handles the provider-specific resume mechanics internally.

The MCP tool is also easier to compose into skills. /consult, /collab, and /debate all just say "call consult_llm with these parameters." A CLI version would need each skill to either teach the agent the CLI's interface or reference a separate skill that does. A skill that orchestrates a multi-model debate is ~90 lines with MCP. As shell commands, the same skill would either balloon into hundreds of lines of escaping rules and stdout parsing, or depend on another skill that teaches the agent how to call each CLI.

If you only need a single provider with simple prompts, a Bash call to gemini or codex with some jq filtering will work fine. MCP starts to make more sense with multiple backends, multi-turn conversations across providers, or custom workflows that nicely compose on top.

Development

To work on the MCP server locally and use your development version:

  1. Clone the repository:

    git clone https://github.com/raine/consult-llm-mcp.git
cd consult-llm-mcp

  2. Build and test:

    cargo build
cargo test
just check  # format, lint, test

  3. Add the MCP server to Claude Code using your local build:

    claude mcp add consult-llm -- /path/to/consult-llm-mcp/target/debug/consult-llm-mcp

Now when you make changes, rebuild with cargo build and restart Claude Code.

Releasing

scripts/publish patch  # or minor, major

This bumps the version in package.json and Cargo.toml, commits, tags, and pushes. GitHub Actions handles cross-compilation and npm publishing.

Related projects

  • workmux — Git worktrees + tmux windows for parallel AI agent workflows
  • claude-history — Search and view Claude Code conversation history with fzf
  • tmux-file-picker — Pop up fzf in tmux to quickly insert file paths, perfect for AI coding assistants

About

MCP server for consulting powerful reasoning models in Claude Code

Resources

Readme

License

MIT license
Activity

Stars

68 stars

Watchers

1 watching

Forks

12 forks
Report repository

Releases 15

v2.13.0 Latest
Mar 29, 2026
+ 14 releases

Packages

 
 
 

Contributors 2

Languages