
consult-plugin
Consult plugin allows you to bridge claude code with codex and agy natively through a comprehensive mcp server.
Install with your AI
Paste into Claude Code, Cursor, or any agent — it reads the repo and wires the tool into your project.
Install and set up consult-plugin (claude-plugin project) into my current project. Found on https://claudeers.com/consult-plugin Repo: https://github.com/NoCritics/consult-plugin Homepage/docs: — Detected install method: claude-plugin → /plugin install consult-plugin@NoCritics/consult-plugin Category: mcp-servers. Platforms: cli, api, desktop. Read the repo's README for exact setup and env vars, then install it and wire it into my project. Claudeers Health Verdict: active; community-verified: false. Confirm the source before running anything.
/plugin marketplace add NoCritics/consult-plugin /plugin install consult-plugin@NoCritics/consult-plugin
git clone https://github.com/NoCritics/consult-plugin
// compatibility
| Platforms | cli, api, desktop |
|---|---|
| Operating systems | — |
| AI compatibility | claude |
| License | MIT |
| Pricing | open-source |
| Language | JavaScript |
consult
Ask other AI models from Claude Code or Claude Desktop: Codex, Antigravity (agy), local models (llama-swap, Ollama, LM Studio, llama-server or any OpenAI-compatible server) and the legacy Gemini CLI.
- Questions and multi-turn chats, with images.
- A council: several models answer, then rank each other's answers.
- Image generation with agy or with GPT through Codex.
- Background tasks for Codex, agy or a local model while you keep working (Claude Code only).
Install
Needs Node.js 20 or newer. In Claude Code:
/plugin marketplace add NoCritics/consult-plugin
/plugin install consult@nocritics
Restart Claude Code. There is nothing to npm install.
Then set up the providers you want. Each one is optional; a tool whose provider is missing tells you what to do.
| Provider | Setup | Used by |
|---|---|---|
| Codex | install the codex CLI, run codex login | questions, chats, council, GPT images, background tasks |
| Antigravity | install agy from antigravity.google, run agy once to sign in | questions, chats, council, images, background tasks |
| Local model | a running runner, then /consult:setlocal | questions, council, background tasks (these also need the pi CLI) |
| Claude | the claude CLI | council |
| Gemini CLI (legacy) | the gemini CLI and a paid GEMINI_API_KEY; the free tier no longer works | ask_gemini |
Installing agy on Windows:
curl -fsSL https://antigravity.google/cli/install.cmd -o install.cmd && install.cmd && del install.cmd
consult finds agy in its install folder even before your PATH picks it up.
Use
Ask in plain words, or use a slash command.
| To | Say or type |
|---|---|
| Get a second opinion | "Get a second opinion on this approach" |
| Ask one model | "Ask codex why …", or /consult:codex, /consult:agy, /consult:local, /consult:gemini |
| Include files | "Ask codex about src/utils.py". Claude puts the file in the prompt, since the other model can't read your disk. |
| Show an image | "Ask codex what's wrong with C:\shots\err.png" |
| Chat | "Start a chat with codex about …", then keep replying, or /consult:chat |
| Run a council | "Run the council on this", or /consult:council |
| Make an image | "Make an image of a lighthouse at dusk". Say "with GPT" for Codex. |
| Hand off a task | /consult:delegate codex <task> |
| Check status | "Is codex signed in?", "Is my local runner up?" (free) |
Time limits. The default is 3 minutes, and at least 10 for Codex. Any call can set its own limit from 30 seconds to 30 minutes: say "give codex 20 minutes", or add --timeout 20m to a slash command. A call that runs out returns what arrived so far, marked INCOMPLETE:.
In Claude Code
- The band. While a call runs, a row above the prompt shows what kind of call it is (
ask,chat,image,council), the model, a bar against its time limit, and Cancel, which stops that call only. A background task gets adelegaterow for as long as it runs, with its time against the task's limit and a Cancel of its own (a task whose fence is lifted leads withunfenced); with only tasks shown, the band's button opens the Delegated tab. A row that ends while others still run stays as its result line (the three newest; older ones fold into one line); when the last one ends, one short row says how it went. Either stays until your next prompt. The rows fit a band narrowed by the/consultpane. - The
/consultpane has three tabs:- Answers: this session's and earlier sessions' answers (the last 50), with copy buttons. Image entries can be opened, shown in their folder or copied as a path. kitty and Ghostty display PNGs in the pane.
- Local models: choose the runner and a model for each nickname. Shows the settings change before you save.
- Delegated: your background tasks, with Cancel, Clear, Discard, their reports and logs.
- The status line checks your providers when a session starts, e.g.
codex ✓ agy ✓ local –. - Notifications: a toast when a call ends, and a push notification for calls that take 2 minutes or more.
Local models
/consult:setlocal finds the runner on this machine and proposes a model for each nickname: default, fast, vision, vision-best and big. It shows you the change and writes only what you confirm.
- llama-swap: uses its own aliases. Give its address once if it isn't on port 9292 or 8080.
- Ollama and LM Studio: proposes models by size, quantization, vision support and whether they fit your GPU.
- llama-server and other OpenAI-compatible servers: model names as they are. Your own server can describe its models by adding a
metaobject to each/v1/modelsentry, e.g.{"arch": "MoE", "engine": "vLLM", "decode_tps": 40, "vision": true}."chat": falsehides a model, such as an embedding model.
Several runners can be set up, and one is in use: /consult:setlocal ollama switches. Use the nicknames anywhere: "ask local fast …", /consult:local --model big ….
Runner settings live in ~/.claude/consult.local.md only; a project's file can't change them. For a key, write the name of the environment variable that holds it (llama-swap.key_env: LLAMA_SWAP_API_KEY). consult reads the variable at call time and sends the key only to this machine.
- Thinking models can spend the whole call thinking.
/consult:local --no-think(reasoning_effort: none) turns it off. Use it when you see "produced only reasoning". - Images cost tokens per pixel, so consult shrinks PNGs to 1024 px on the long edge (
local_image_max_edge). Use a model of 9B parameters or more for images. - Slow models: every llama-swap call, and any Ollama model that doesn't fit your GPU, gets at least 20 minutes.
Background tasks
/consult:delegate <worker> <task> hands work to Codex, agy or a local model. It runs in the background. When it ends, Claude is woken up, reviews the result and merges it. Claude Code only.
-
Workers:
codex,agy,localorlocal:<nickname>;gpt,geminiandqwenwork too. With no worker named, Claude picks one by the kind of task and tells you which:- code goes to Codex;
- research and reviews go to agy;
- private work goes to a local model.
Change these in the Delegated tab's settings.
-
Fencing: each project has a fence per worker, set in the Delegated tab's settings. Codex is fenced by its own sandbox and stays inside its folder; set to full access, its write tasks can change any file and run any command on this PC (read tasks still change nothing, but may read anywhere). agy and the local worker (the
piagent) aren't fenced by anything: by default their brief asks them to stay in the folder, and anywhere drops that. agy can't run commands. Changes a worker makes outside the project don't go through the copy and can't be undone, and nothing stops an unfenced worker from editing consult's own settings; a stronger fence is planned. A task keeps the fence it started with: every run uses the stricter of that and the project's current setting, so lifting a fence never widens a task already started, while putting it back narrows it on the next resume. -
Copies: in a git project (git 2.40 or newer, at least one commit), a task that edits files works in a separate copy that includes your uncommitted changes.
- Claude merges the copy back file by file. A file changed on both sides gets conflict markers.
- Nothing is committed, and your branch and staged changes are left alone.
- Files that configure tools (dot-files,
CLAUDE.md,AGENTS.md, lint configs) are saved as<file>.taskuntil you approve them. - Without git, tasks edit in place. Claude offers once to set git up.
-
Usage limits: consult reads Codex's usage and asks before starting at 90 % or more. A task that hits a limit pauses and can continue after the reset.
-
Local tasks never swap out a model another task is using.
-
Stopping:
- Tasks stop when Claude Code closes, unless you set them to keep running.
- Each has a time limit (3 hours by default) and is stopped after 10 minutes with no output.
- When a task ends, consult stops the processes the worker started. On Windows it also stops a dev server that a task in a copy left running. For a task in your own folder it only warns you, since it can't tell that server from yours.
-
Storage: task data lives in
~/.claude/consult-delegate/. consult never commits, stashes or resets in your project.
Claude Desktop
The server also installs as a Claude Desktop extension, for Chat and Cowork.
- Install: download
consult-<version>.mcpbfrom the releases page, then go to Settings → Extensions → Advanced → Install extension. Don't use Customize → Plugins: that adds a second copy to Claude Code. - Requirements: leave Use built-in Node.js for MCP on. codex and agy must be installed and signed in.
- What's included: Desktop gets 17 tools, but no slash commands, band, pane or background tasks. Ask for a model by name ("ask codex …").
- Settings come from
~/.claude/consult.local.md. - Cowork stops waiting for a tool after 4 minutes, and consult can't change that. For quick answers there, ask for low effort.
- Updating: install the new file and start a new chat.
Images
- Image tools:
agy_generate_imageis the default.codex_generate_imageis used when you name GPT, OpenAI, ChatGPT or Codex, or need a transparent background. - Where they're saved:
consult-images/in the project, else~/Pictures/consult-images, or the folder set inagy_image_dirorcodex_image_dir. Existing files are never overwritten. - Options:
- aspect ratio;
- 1–4 images;
- your prompt word for word;
- reference images (agy up to 3, Codex up to 5);
- a transparent background (Codex only).
- agy: JPEG at about 1024 px, about a minute per image.
- Codex: PNG, about 45 seconds per image. It uses your ChatGPT plan's Codex quota; no API key needed.
- Reading images:
ask_codex,ask_agy,ask_localand the council take image files (PNG, JPEG, GIF, WebP).
Council
/consult:council has Claude, Codex and agy each answer, then each ranks all the answers without knowing who wrote which. Points are added up (Borda count) and the highest wins.
- If one model fails, the council continues with two. A tie is shown to you to pick.
- Gemini and local models take part only when named:
--participants claude,codex,agy,local. - The Claude participant runs without your
CLAUDE.md, plugins or MCP servers, which keeps it cheap. - With images, only the answering round sees them, and a council that includes Claude or Gemini refuses them.
Settings
Optional. They go in ~/.claude/consult.local.md (global) or <project>/.claude/consult.local.md.
- Which file is used: the project file when it exists, else the global one.
- Global-only settings: runner and background-task settings come from the global file only.
- Every key: examples/consult.local.md lists each one with its default.
- Mistakes:
local_statusreports unknown keys and lines it can't read.
Good to know
- Antigravity can write files, and consult can't stop it.
- agy creates and edits files without asking in any folder listed in its
trustedWorkspaces(~/.gemini/antigravity-cli/settings.json). - consult tells it not to and reports what it tried.
agy_statusshows your setting. - To block writes, add this to that file:
"permissions": { "deny": ["write_file(**)", "command(**)"] }
- agy creates and edits files without asking in any folder listed in its
- Codex:
- It is read-only by default (
codex_sandbox), and a one-off question leaves no saved session. - Your
~/.codex/hooks.jsondoesn't run on consult calls unless you ask, or setcodex_hooks: true. codex_login_statuscan report "Logged in" after the sign-in has expired. If calls fail with a token error, runcodex login.
- It is read-only by default (
- Cost:
- Every call uses tokens on its provider, except local models.
- agy uses about 20,000–29,000 input tokens per call, even for a short question.
- Cleanup:
- Cancelling a call stops the CLI behind it.
- agy's saved conversation is deleted after a one-off question. A chat keeps its conversation until you end the chat.
- Long answers: in Claude Code, an answer over 40,000 bytes is saved in full to
~/.claude/consult-answers/, which keeps the last 100. Claude gets the first 8,000 characters and the file's path. - Remote servers: no tool takes a server address, so a prompt can't send your prompts elsewhere. consult warns if a project's settings point a local runner at another machine.
More
- CHANGELOG.md lists what changed in each version.
- CONTRIBUTING.md covers the tests, the file layout and design notes.
- Tests run on Linux, macOS and Windows for every push.
// faq
What is consult-plugin?
Consult plugin allows you to bridge claude code with codex and agy natively through a comprehensive mcp server.. It is open-source on GitHub.
Is consult-plugin free to use?
consult-plugin is open-source under the MIT license, so it is free to use.
What category does consult-plugin belong to?
consult-plugin is listed under mcp-servers in the Claudeers registry of Claude-compatible tools.
// embed badge
[](https://claudeers.com/consult-plugin)
// retro hit counter
[](https://claudeers.com/consult-plugin)
// reviews
// guestbook
// related in MCP Servers
f.k.a. Awesome ChatGPT Prompts. Share, discover, and collect prompts from the community. Free and open source — self-host for your organization with complete…
A cross-platform desktop All-in-One assistant for Claude Code, Codex, OpenCode, OpenClaw, Gemini CLI & Hermes Agent. Only official website: ccswitch.io
🪨 why use many token when few token do trick — Claude Code skill that cuts 65% of tokens by talking like caveman
An open-source AI agent that brings the power of Gemini directly into your terminal.