CommunityActive· 6d★ 4.3by amidabuddha

Unichat MCP

Sends requests to OpenAI, Anthropic, MistralAI, xAI, Google AI, or DeepSeek through tool calls or predefined prompts.

View repo

Safety & Trust

Delv Safety Grade: C

Score 61/100 · assessed 2026-04-28

Maintainer45

Permissions65

Supply chain70

Transparency75

Incidents100

Unichat MCP is a community-maintained proxy that routes LLM requests from Claude to six different AI providers. The maintainer (amidabuddha) appears to be a solo developer with moderate GitHub activity. The server requires multiple API keys as environment variables, creating a broad credential surface. Permissions are moderately scoped: it makes outbound network calls to third-party LLM APIs and reads secrets from environment variables, but doesn't touch the filesystem or execute shell commands. Supply chain is reasonable via PyPI (uvx install), though the package has limited adoption signals. Transparency is decent with open source code and basic documentation. The core risk is credential exposure: you're handing API keys for up to six providers to a single community tool. No known security incidents, but the bus factor is high and the attack surface spans multiple vendor APIs. Suitable for experimentation but requires careful key management for production use.

Lethal Trifecta (prompt-injection exposure)

ONE OF THREE

Private dataNo

Reads secrets, credentials, private files

Untrusted inputNo

Ingests web pages, PRs, issues, emails

External commsYes

Can send data outbound

Generic LLM chat passthrough.

Green flags

Open source with visible code on GitHub for audit
Distributed via PyPI with standard uvx installation
Scoped to API calls only, no filesystem or shell access
No known security incidents or CVEs
Clear use case for model comparison and evaluation workflows

Red flags

Solo maintainer with limited community review or adoption signals
Requires API keys for up to six different LLM providers simultaneously
High blast radius if credentials leak: multi-vendor API key exposure
Limited maintenance history and unclear long-term support commitment

Permissions requested

Outbound networkAccess secretsExternal LLM call

Assessed by Delv Editorial using public metadata. Grades are advisory and update as the ecosystem changes. They do not replace your own review of permissions and code before granting an agent access to sensitive systems.

Install

uvx unichat-mcp-server

Env vars needed: OPENAI_API_KEY

Review

Unichat MCP is a proxy server that lets Claude call other LLMs during a conversation. You configure API keys for OpenAI, Anthropic, MistralAI, xAI, Google AI, or DeepSeek, then Claude can send prompts to any of them mid-task. It's not a chat interface. It's a tool that turns Claude into a router. I'd reach for this when I'm building evals or need to compare model outputs without switching tabs. Say you're testing prompt variations: Claude can fire the same prompt at GPT-4o, Gemini, and DeepSeek in one go, then summarise the differences. Or you're prototyping a multi-agent system where one model generates code and another reviews it. Unichat makes that a single MCP call instead of duct-taping API clients together. The repo supports predefined prompts, which means you can template common requests and invoke them by name. Useful if you're running the same eval questions across models repeatedly. The tool calls are straightforward: you specify the provider, model, and prompt. Claude handles the rest. Quirks: you need API keys for every provider you want to use, and the server doesn't validate them upfront. If a key is wrong, you'll find out when Claude tries to call it. The documentation is light on examples, so expect to read the code if you want to do anything beyond basic calls. No streaming support, which matters if you're calling models that take a while to respond. And it's Python-based, so startup time is noticeable if you're restarting Claude Desktop often. This isn't for casual users. If you just want to try different models, open their web UIs. But if you're building workflows where Claude needs to orchestrate other LLMs, or you're running systematic comparisons, Unichat is the fastest way to wire it up. It does one thing and stays out of the way.

Verdict

Install this if you're building evals, multi-model agents, or need Claude to call other LLMs as part of a workflow. Skip it if you're just curious about other models or don't have API credits to spare. It's a developer tool, not a convenience feature.

Good at

Lets Claude call six major LLM providers without leaving the conversation.
Predefined prompts make it easy to run repeatable evals across models.
Minimal setup if you already have API keys lying around.
Useful for multi-agent workflows where different models handle different tasks.

Watch out

No streaming, so long responses from other models will hang visibly.
API key errors only surface when Claude tries to use them, not at startup.
Documentation is sparse. You'll need to read the source for anything non-trivial.
Requires active API credits for every provider you want to route to.

Use cases

cross-model comparison
fallback routing
eval harnesses
multi-provider agents

Getting started

1. Run `uvx unichat-mcp-server` to install. 2. Add API keys for the providers you want (OPENAI_API_KEY, ANTHROPIC_API_KEY, etc.) to your environment or Claude Desktop config. 3. Add the server to `claude_desktop_config.json` under `mcpServers` with `command: "uvx"` and `args: ["unichat-mcp-server"]`. 4. Restart Claude Desktop and ask it to call a model via Unichat (e.g., "Use Unichat to ask GPT-4o what 2+2 is"). 5. Watch out for missing API keys. The server won't warn you until Claude tries to use a provider you haven't configured.

Works with

Claude DesktopClaude CodeCursor