Skip to main content
Glama
README.md
# mosaic-mcp

<!-- mcp-name: io.github.sourabhnk/mosaic-mcp -->

Coverage-honest pre-clinical target intelligence over MCP. Every count carries
a coverage state, so "we did not look" never reads as zero.

**3.0.0 is a thin client to the hosted Mosaic service.** It serves Mosaic's
five MCP tools over stdio and forwards every call to
`https://mcp.getmosaic.dev/mcp` with your API key. It holds no database and no
query layer: the answers, and any refusal, are the hosted service's.

> **Deprecated: the 2.x bring-your-own-database mode.** Versions up to 2.2.1
> ran Mosaic's query layer over a PostgreSQL database you provided
> (`DATABASE_URL`). That mode is deprecated and is not in 3.0.0; `DATABASE_URL`
> is ignored. Its eleven tools answer an old client by name with what replaced
> them. To keep running it, pin `mosaic-mcp<3` (2.2.1 is the last release).

## Do you need this package?

If your MCP client can connect to a remote server, you do not: add
`https://mcp.getmosaic.dev/mcp` as a connector (in claude.ai: Settings →
Connectors → Add custom connector). It signs you in by email (OAuth) and needs
no key and no install.

This package is for clients that only run local (stdio) servers.

## Quick start

```bash
pip install mosaic-mcp          # or: uvx mosaic-mcp
export MOSAIC_API_KEY="mk_live_..."
mosaic-mcp                      # stdio MCP server
```

### With Claude Desktop

```json
{
  "mcpServers": {
    "mosaic": {
      "command": "uvx",
      "args": ["mosaic-mcp"],
      "env": { "MOSAIC_API_KEY": "mk_live_..." }
    }
  }
}
```

### With Claude Code

```bash
claude mcp add mosaic --env MOSAIC_API_KEY=mk_live_... -- uvx mosaic-mcp
```

## Tools

| Tool | Tier | What it does |
|------|------|--------------|
| `mosaic_get_coverage_grid` | free | A coverage grid for any human gene symbol: every axis with its state; network axes `queued` until a dossier fetches them |
| `mosaic_verify_claim` | free, rate-limited | One claim about a target checked against Mosaic's own independent count, in at most 10 s, with one of seven verdicts (`confirmed`, `contradicted`, `cannot_verify_same_source`, `cannot_verify_floor`, `cannot_verify_unavailable`, `cannot_verify_unscoped`, `out_of_scope`) |
| `mosaic_audit_claims` | paid ($390 an audit) | Every target-level claim in a document (a press release, a filing section, an AI-generated assessment), audited; async |
| `mosaic_request_dossier` | paid ($149 a dossier) | A verified dossier for any human target; async |
| `mosaic_get_result` | free | The status and result of a dossier (`dossier-N`) or audit (`audit-N`) you requested; only your own |

The tier, the price and any refusal come from the hosted service; this package
decides none of them.

## Configuration

| Variable | Required | Description |
|----------|----------|-------------|
| `MOSAIC_API_KEY` | Yes | Your Mosaic API key (`mk_live_…`), sent as a Bearer token. Without it every call says so and sends nothing. |
| `MOSAIC_MCP_URL` | No | The hosted MCP endpoint. Default `https://mcp.getmosaic.dev/mcp`. |

A call the hosted service could not answer (unreachable, a timeout, a rejected
key) is returned as an error saying so: never as an empty or a zero result.

## Changelog

Full history in [CHANGELOG.md](CHANGELOG.md). **3.0.0**: a thin client to the
hosted service with the five tools of plan v3 §3.1; the bring-your-own-database
mode and its eleven tools are retired (each answers with its replacement).

## License

Apache 2.0.

TDQS

A4/5.0

Scored across 44 tools

Disambiguation5/5

Each tool has a clearly distinct purpose, targeting specific aspects of drug discovery like target assessment, compound analysis, competitive landscape, and clinical trials. The descriptions are detailed enough to avoid confusion, even for similar-sounding tools.

Naming Consistency5/5

All tools follow the 'mosaic_<verb>_<noun>' or 'mosaic_<noun>_<verb>' pattern in snake_case, with verbs like 'get', 'compare', 'find', 'list', and 'assess'. The naming is predictable and uniform, aiding tool selection.

Tool Count4/5

With 44 tools, the server covers a broad domain thoroughly. While the count is high, each tool serves a specific function without redundancy, and the scope justifies the number. A slight reduction could improve manageability, but it's still well-scoped.

Completeness5/5

The tool set comprehensively covers drug discovery workflows: target discovery, druggability assessment, compound profiling, competitive intelligence, clinical trials, and more. Gaps are acknowledged and minimal, making the surface complete for its stated purpose.

Maintenance

ActivityActive
ResponsivenessNo recent community issues