Vox MCP is a multi-model AI gateway that lets MCP clients send prompts to any configured external AI model, with conversation memory, file/image context, and thread export.
Chat with any configured provider — route prompts to Gemini, OpenAI, Anthropic, xAI, DeepSeek, Moonshot, OpenRouter, Cloudflare/Vercel AI Gateways, or a custom endpoint (Ollama, vLLM, LM Studio).
Get raw, unmodified responses — no system prompt injection, formatting, or behavioral directives; pure passthrough plus routing.
Attach context — pass absolute file paths (files or directories, expanded recursively) and images (absolute paths or base64) alongside the prompt.
Control generation — set temperature (0–2, clamped per model) and thinking_mode (minimal → max) for reasoning depth.
Hold multi-turn conversations — reuse a continuation_id to keep history and files across turns, even switching providers mid-thread (start on Gemini, continue on GPT).
Let the agent pick a model — omit/use auto and the calling agent selects from available options; explicit models and provider defaults take precedence.
Discover available models — call listmodels to see configured providers, model names, aliases, and capabilities.
Export threads — call dump_threads to export conversations as Markdown (with YAML frontmatter, written to disk) or raw inline JSON, optionally filtered by thread UUIDs.
Rely on persistence — threads are stored in memory and shadow-persisted to JSONL, so they can be cold-reloaded after expiry (default TTL 24h, 100 turns).
Allows sending prompts to Google Gemini models (e.g., gemini-2.5-pro) via the Gemini API, with optional file and image context.
Allows sending prompts to local Ollama models by configuring a custom API endpoint.
Allows sending prompts to OpenAI models (e.g., GPT-5, o3, o4-mini) via the OpenAI API.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@Vox MCP chat with gemini: what is dark matter?"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Multi-model AI gateway for MCP clients.
MCP clients like Claude Code, Claude Desktop, and Cursor are locked to their host model. Vox gives them access to every other model — Gemini, GPT, Grok, DeepSeek, Kimi, or your local Ollama — through a single chat tool.
The design is deliberately minimal: prompts go to providers unmodified, responses come back unmodified. No system prompt injection. No response formatting. No behavioral directives. The only value Vox adds is routing and conversation memory — everything else is pure passthrough.
Related MCP server: 1mcpserver
Send a prompt, optionally attach files or images, pick a model (or let the agent pick), and get back the model's raw response. Conversation threads persist in memory via continuation_id for multi-turn exchanges across any provider — start a thread with Gemini, continue it with GPT. Threads are shadow-persisted to disk as JSONL for durability and can be exported as Markdown.
3 tools:
Tool | Description |
| Send prompts to any configured AI model with optional file/image context |
| Show available models, aliases, and capabilities |
| Export conversation threads as JSON or Markdown |
10 providers:
Provider | Environment variable | Built-in preference or route |
Google Gemini |
|
|
OpenAI |
|
|
Anthropic |
|
|
xAI |
|
|
DeepSeek |
|
|
Moonshot (Kimi) |
|
|
OpenRouter |
| Explicit OpenRouter model ID or curated alias |
Cloudflare AI Gateway |
|
|
Vercel AI Gateway |
|
|
Custom |
| Ollama, vLLM, LM Studio, etc. |
These are preferences within each provider. With auto, the calling agent chooses
an available model; if Vox must choose, its existing provider priority applies.
An explicit model or configured default takes precedence. GPT-6 Sol/Luna, Grok 4.7,
Claude Fable 5.1, Sonnet 5, and Haiku 4.5 remain selectable. Claude 3 Opus's pinned
snapshot is preserved for accounts with access.
See model selections and verification sources for provider preferences, exact API IDs, and reasoning behavior.
With uv installed, add Vox to your MCP client's configuration. For example, to use OpenAI:
{
"mcpServers": {
"vox-mcp": {
"command": "uvx",
"args": ["vox-mcp"],
"env": {
"OPENAI_API_KEY": "your-key-here"
}
}
}
}Use the environment variable for your preferred provider from the table above.
Only one provider is required. Connect the server, then ask your agent to call
listmodels to see the models available with your configuration.
For an exact release, use "args": ["vox-mcp@0.8.0"]. To refresh a cached
installation, run uvx --refresh vox-mcp config show, then reconnect the MCP server.
A version pinned in your client configuration must be updated there as well.
git clone https://github.com/linxule/vox-mcp.git
cd vox-mcp
cp .env.example .env
# Edit .env — add at least one API key
uv sync
uv run python server.pyVox runs as a stdio MCP server. Each client needs to know how to launch it.
The following examples use a source checkout; the published-package configuration
above avoids cloning the repository. Replace /path/to/vox-mcp with the absolute
path to your cloned repo.
claude mcp add vox-mcp \
-e GEMINI_API_KEY=your-key-here \
-- uv run --directory /path/to/vox-mcp python server.pyOr add to .mcp.json in your project root:
{
"mcpServers": {
"vox-mcp": {
"command": "uv",
"args": ["run", "--directory", "/path/to/vox-mcp", "python", "server.py"],
"env": {
"GEMINI_API_KEY": "your-key-here"
}
}
}
}Add to claude_desktop_config.json:
macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
Windows: %APPDATA%\Claude\claude_desktop_config.json
{
"mcpServers": {
"vox-mcp": {
"command": "uv",
"args": ["run", "--directory", "/path/to/vox-mcp", "python", "server.py"],
"env": {
"GEMINI_API_KEY": "your-key-here"
}
}
}
}Add to .cursor/mcp.json (project) or ~/.cursor/mcp.json (global):
{
"mcpServers": {
"vox-mcp": {
"command": "uv",
"args": ["run", "--directory", "/path/to/vox-mcp", "python", "server.py"],
"env": {
"GEMINI_API_KEY": "your-key-here"
}
}
}
}Add to ~/.codeium/windsurf/mcp_config.json:
{
"mcpServers": {
"vox-mcp": {
"command": "uv",
"args": ["run", "--directory", "/path/to/vox-mcp", "python", "server.py"],
"env": {
"GEMINI_API_KEY": "your-key-here"
}
}
}
}The canonical stdio configuration:
{
"mcpServers": {
"vox-mcp": {
"command": "uv",
"args": ["run", "--directory", "/path/to/vox-mcp", "python", "server.py"],
"env": {
"GEMINI_API_KEY": "your-key-here"
}
}
}
}Tips:
Paths must be absolute
You only need one API key to start — add more providers later via .env
The .env file in the vox-mcp directory is loaded automatically, so API keys can go there instead of in the client config
Use VOX_FORCE_ENV_OVERRIDE=true in .env if client-passed env vars conflict with your .env values
Set these values in your MCP client's Vox env object. For a source checkout,
you can instead copy .env.example to .env:
API keys — at least one provider key is required
DEFAULT_MODEL — overrides the saved model preference; auto asks the agent to choose
VOX_CONFIG_PATH — optional settings file path (default: ~/.vox/config.json)
Model restrictions — GOOGLE_ALLOWED_MODELS, OPENAI_ALLOWED_MODELS, etc.
CONVERSATION_TIMEOUT_HOURS — thread TTL (default: 24h)
MAX_CONVERSATION_TURNS — thread length limit (default: 100)
See .env.example for the full reference.
Available since 0.8.0. You or your agent can save a default without editing MCP client configuration. For example, choose GPT-6 Astra:
uvx vox-mcp config set-default gpt-6-astra
uvx vox-mcp config showReplace gpt-6-astra with any available model ID, such as kimi-k3, grok-4.6,
claude-opus-5-5, claude-fable-5-1, deepseek-flash, or gemini-3.8-flash.
To remove the saved preference, run uvx vox-mcp config reset-default.
For a source checkout, use uv run vox-mcp config .... Preferences are stored in
~/.vox/config.json; set VOX_CONFIG_PATH to use another file. This file contains
model preferences, not API keys. Commands make no model requests and do not need
provider credentials. Use listmodels in your connected MCP client to choose an
available ID or alias. Explicit gateway routes such as
vercel/google/gemini-3.1-pro-preview can also be saved. IDs are preserved exactly;
saving a default does not verify your account's access to that model.
Restart or reconnect Vox after changing settings. config show prints the saved
default, effective default, and its source for the command's launch environment.
The running MCP client's environment may differ; listmodels reports the default
that its server actually loaded.
Selection order is: explicit chat.model, the previous model for a continued
thread, DEFAULT_MODEL from the server environment, the saved default, then auto.
A saved default therefore changes new conversations; it does not switch existing
threads. To override saved preferences for one MCP client, add
"DEFAULT_MODEL": "your-model-id" to the Vox server's env object. Existing source
checkout .env settings also retain their environment precedence. The automatically
loaded .env belongs to the Vox installation, not the current directory; use the
settings command or MCP client environment for uvx installations.
With auto, the agent is asked to select a model for each call. If a caller passes
model: "auto", Vox uses its built-in provider priority and preferences; this is
not a live comparison of all providers. Explicit model selections always remain
available. An agent with terminal access can do this for you:
Use Vox listmodels to check that GPT-6 Astra is available. Save it as my Vox default with the config command, check whether DEFAULT_MODEL overrides it, and tell me how to reconnect Vox.
An agent with only Vox's three MCP tools can select a model for each call, but cannot persist a preference; saving settings requires terminal or file access.
Use an explicit gateway prefix in chat.model. Vox removes only that first prefix
before sending the request and keeps it in conversation memory. A missing gateway
configuration or disallowed model fails without falling through to OpenRouter or a
native provider. Bare model names keep their existing routing behavior.
CLOUDFLARE_API_TOKEN=your-cloudflare-token
CLOUDFLARE_ACCOUNT_ID=your-account-id
CLOUDFLARE_GATEWAY_ID=default
CLOUDFLARE_MODELS=openai/gpt-5.5{"prompt": "Explain quorum consensus.", "model": "cloudflare/openai/gpt-5.5"}Vox uses the Cloudflare account REST API
at https://api.cloudflare.com/client/v4/accounts/<account>/ai/v1, authenticates
with a bearer token, and sets cf-aig-gateway-id (default unless configured).
The token needs Workers AI Read permission; an AI Gateway-only token is not
sufficient. Third-party models use Cloudflare Unified Billing. Workers AI model
IDs retain their @cf/ prefix, for example cloudflare/@cf/moonshotai/kimi-k2.6.
Legacy /compat, provider-key forwarding, and dynamic/ routes are not supported
by this adapter.
VERCEL_AI_GATEWAY_API_KEY=your-vercel-gateway-key
VERCEL_MODELS=anthropic/claude-sonnet-4.6{"prompt": "Explain quorum consensus.", "model": "vercel/anthropic/claude-sonnet-4.6"}Vox uses Vercel's OpenAI-compatible API
at https://ai-gateway.vercel.sh/v1. AI_GATEWAY_API_KEY is also accepted;
VERCEL_AI_GATEWAY_API_KEY takes precedence when both are set.
CLOUDFLARE_MODELS and VERCEL_MODELS are optional comma-separated upstream IDs
for listmodels and agent discovery. They do not restrict access. Explicit gateway
model IDs work without a catalog, including with the default DEFAULT_MODEL=auto;
the caller must supply the gateway model. Use CLOUDFLARE_ALLOWED_MODELS or
VERCEL_ALLOWED_MODELS to restrict access. Both upstream IDs and fully prefixed
Vox routes are accepted in catalogs and allowlists. Use the exact model ID
published by the gateway; native-provider and gateway IDs can differ.
Gateway requests currently support text only. Vox does not infer vision or
thinking controls from a model name; explicit gateway thinking_mode requests are
rejected before inference. Its 32,768-token context and 4,096-token output
budgets are conservative local estimates, not advertised upstream limits; these
numbers are not sent as generation parameters. Omitted temperature and reasoning
settings use upstream defaults. No catalog or model availability request is made
at startup. The adapters are covered by mocked HTTP tests; live inference requires
a configured account and has not been exercised as part of the release checks.
Dependencies are maintained in pyproject.toml and uv.lock; Dependabot updates
the lock through its uv integration while respecting the supported version ranges.
Changing those ranges requires a deliberate compatibility review. CI checks the lockfile, runs the offline test
suite on Python 3.10 and 3.13, and audits locked packages with pip-audit.
The supported SDK lines are MCP 1.x, OpenAI 2.x, and Anthropic 0.x.
MCP SDK 2 requires a separate server API migration; provider major upgrades
are kept separate from dependency maintenance.
uv sync
uv run python -c "import server" # smoke test
uv run pytest # run testsSee CONTRIBUTING.md for code style, project structure, and how to add providers.
Apache 2.0 — see LICENSE and NOTICE.
Derived from pal-mcp-server by Beehive Innovations.
Multi-model AI gateway. Routes prompts to external AI models (Gemini, OpenAI, Anthropic, DeepSeek, Moonshot, xAI, OpenRouter, custom endpoints) with conversation memory. Supports file context embedding, images, and multi-turn threads via continuation_id.
| Name | Required | Description | Default |
|---|---|---|---|
| model | Yes | Currently in auto model selection mode. If no model is provided, you may use the `listmodels` tool to review options and select an appropriate match. The server validates model availability and returns errors for unknown models. Top models: gemini-3.8-flash (score 100, 1.0M ctx, thinking, code-gen); gemini-2.5-pro (score 100, 1.0M ctx, thinking, code-gen); gemini-3.1-pro-preview (score 100, 1.0M ctx, thinking, code-gen); gemini-2.5-flash (score 81, 1.0M ctx, thinking). | |
| images | No | Image paths (absolute) or base64 strings for optional visual context. | |
| prompt | Yes | Your question or task for the external model. Prefer passing code and large content via absolute_file_paths rather than inlining it here. | |
| temperature | No | Optional sampling temperature. If omitted, the model's own default is used (recommended; some reasoning models reject or degrade on a fabricated value). Range is provider-dependent (commonly 0–2); values are clamped per model. | |
| thinking_mode | No | Optional reasoning depth: minimal, low, medium, high, or max. Omit to use the provider default. | |
| continuation_id | No | Unique thread continuation ID for multi-turn conversations. Works across different tools. Reuse the last continuation_id you were given to preserve full conversation context, files, and history across turns. Threads are held in memory and expire after inactivity. | |
| absolute_file_paths | No | Full, absolute file paths to relevant code in order to share with the external model. Accepts both files and directories (directories are expanded recursively). Content is read and embedded into the prompt context. |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
Annotations already declare readOnlyHint=true, so the safety profile is covered; the description adds real behavioral context beyond that, namely that it is a multi-provider gateway with conversation memory, file-context embedding, image support, and multi-turn threads. It omits any mention of cost, rate limits, or per-provider failure behavior, so it falls short of a 5.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
Three tight sentences, front-loaded with the core identity ('Multi-model AI gateway') followed by routing scope and the capabilities that matter for invocation. No sentence is wasted or redundant with the schema.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a 7-parameter, multi-provider routing tool with no output schema, the description covers the essential mental model: what it routes to, that memory/threads exist, and that files and images can be attached. It leaves unaddressed what happens on provider failure or how responses are shaped, but the readOnly annotation and rich schema compensate.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema description coverage is 100%, so every parameter is already documented in detail, which sets the baseline at 3. The description adds only high-level framing ('file context embedding, images, and multi-turn threads via continuation_id') without new syntax or format detail beyond the schema.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description states a specific verb and resource ('Routes prompts to external AI models') and enumerates the providers, so an agent immediately knows this is a completion/routing tool rather than a listing tool. It does not explicitly name the siblings (listmodels, dump_threads) to contrast itself against them, which keeps it at a 4 rather than a 5.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
Usage is implied by 'Multi-model AI gateway' and the routing sentence, but there is no explicit statement of when to use chat versus listmodels (which the schema, not the description, suggests for model discovery) or when a thread should be continued versus started fresh. The agent must infer the workflow.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Export conversation threads as JSON or Markdown. Threads persist to disk and can be cold-reloaded after memory expiry. Use thread_ids to filter specific threads, format to choose output.
| Name | Required | Description | Default |
|---|---|---|---|
| format | No | Output format: 'markdown' (clean export with YAML frontmatter, written to disk) or 'json' (raw thread data, inline). | markdown |
| thread_ids | No | Filter to specific thread UUIDs. Omit for all active threads. |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
The annotation already declares readOnlyHint=true, so the description only needs to add extra behavioral nuance. It does mention thread persistence and cold-reload, which is useful context beyond the annotation, but it does not clarify whether the export writes files to disk (only the schema does for markdown) or describe any side effects. No contradiction with annotations.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
Two sentences, with the primary purpose front-loaded. The second sentence is a bit of a run-on ('Use thread_ids to filter specific threads, format to choose output') but is still concise and informative. No wasted words.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a simple two-parameter tool with no required fields and no output schema, the description plus schema covers the main usage: exporting threads with optional filtering and format selection. The persistence/cold-reload note adds valuable context. Could mention the default behavior (all threads) but that is already in the schema.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema description coverage is 100%, with both parameters (format and thread_ids) already documented in detail. The description merely echoes 'Use thread_ids to filter specific threads, format to choose output' without adding new semantic information. Baseline 3 applies since schema carries the load.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description starts with a specific verb and resource: 'Export conversation threads as JSON or Markdown.' It clearly identifies the tool's function and distinguishes it from sibling tools like chat and listmodels, which involve interaction and model listing respectively.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description provides implicit usage context: 'Threads persist to disk and can be cold-reloaded after memory expiry' suggests using this tool for backup or recovery after memory loss. It does not explicitly name alternatives or exclusions, but the context is clear enough for an agent to infer when this tool is appropriate.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Shows which AI model providers are configured, available model names, their aliases and capabilities.
| Name | Required | Description | Default |
|---|---|---|---|
No parameters | |||
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
Annotations already declare readOnlyHint=true, so the safe-read nature is covered. The description adds useful context about what information is returned (providers, names, aliases, capabilities), but it does not disclose additional behavioral traits such as pagination or formatting.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is a single, concise sentence that front-loads the verb 'Shows' and packs all relevant information about the tool's output without unnecessary words.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
The tool is simple with no parameters and no output schema, but the description sufficiently covers its purpose and the categories of data it returns. It could be slightly more complete by mentioning that no arguments are required, but that is implicit in the empty input schema.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
The tool has zero parameters, so there is nothing for the description to elaborate. The baseline score of 4 applies, as no parameter information is needed.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description uses a specific verb 'Shows' and clearly defines the resource: configured AI model providers, available model names, aliases, and capabilities. This clearly distinguishes it from sibling tools like chat and dump_threads.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description implies usage—you'd use this tool when you need to see configured models—but it does not provide explicit guidance on when to use it versus alternatives like chat or dump_threads. There is no direct comparison or exclusion.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
v0.8.0chat2 fields changedInput schema / properties / model / descriptionPrevious value: -"Currently in auto model selection mode. If no model is provided, you may use the `listmodels` tool to review options and select an appropriate match. The server validates model availability and returns errors for unknown models. Top models: gemini-2.5-pro (score 100, 1.0M ctx, thinking, code-gen); gemini-3.1-pro-preview (score 100, 1.0M ctx, thinking, code-gen); gemini-2.5-flash (score 81, 1.0M ctx, thinking); gemini-2.0-flash (score 66, 1.0M ctx); gemini-2.0-flash-lite (score 56, 1.0M ctx)."New value: +"Currently in auto model selection mode. If no model is provided, you may use the `listmodels` tool to review options and select an appropriate match. The server validates model availability and returns errors for unknown models. Top models: gemini-3.8-flash (score 100, 1.0M ctx, thinking, code-gen); gemini-2.5-pro (score 100, 1.0M ctx, thinking, code-gen); gemini-3.1-pro-preview (score 100, 1.0M ctx, thinking, code-gen); gemini-2.5-flash (score 81, 1.0M ctx, thinking)."Input schema / properties / thinking_mode / descriptionPrevious value: -"Reasoning depth: minimal, low, medium, high, or max."New value: +"Optional reasoning depth: minimal, low, medium, high, or max. Omit to use the provider default."v0.5.0chatdump_threadslistmodelsScored across 3 tools
chat, dump_threads, and listmodels have clearly distinct purposes: sending prompts, exporting conversation threads, and listing available models. There is no overlap or ambiguity in which tool to use for each task.
The set mixes naming conventions: 'chat' is a single lowercase word, 'dump_threads' uses snake_case verb_noun, and 'listmodels' is a concatenated lowercase noun phrase. While still readable, the pattern is inconsistent.
Three tools is lean but reasonable for a focused multi-model gateway covering chat, export, and model discovery. It could benefit from one or two additional thread-management tools, but nothing feels excessive.
Core operations are present: chatting, exporting threads, and listing models. However, obvious lifecycle operations like deleting or clearing threads are missing, and there is no way to list threads independently of exporting them.
Real-time chat hub for AI agents — Claude Code, Cursor, Cline, Codex over MCP or REST.
One memory, every AI: Claude, ChatGPT, Perplexity, Gemini, Cursor, OpenClaw, Hermes, any MCP client.
One MCP endpoint for Claude, GPT & Gemini: 100+ tools + no-code connectors + agent workers.
Talk to your public-facing AI from any MCP client — Claude, ChatGPT, Cursor, Cline, Windsurf.
