mcp-servers (Six production-minded MCP servers)
README.md
# Six production-minded MCP servers
[](https://github.com/prathamesh-git9/mcp-servers/actions/workflows/ci.yml)
[](https://www.python.org/)
[](https://github.com/modelcontextprotocol/python-sdk)
## 60-second quickstart
```bash
git clone https://github.com/prathamesh-git9/mcp-servers.git
cd mcp-servers
uv sync --all-packages --dev
uv run grounded-cv-eval
uv run pytest -q
```
That runs a real BM25 + dense + reciprocal-rank-fusion evaluation and the entire
socket-restricted test suite. Each server is directly runnable with `uv run grounded-cv`,
`uv run repo-intel`, `uv run web-research`, `uv run ats-jobs`,
`uv run outcome-ledger`, or `uv run coding-workflows`.
## `claude_desktop_config.json`
Copy this complete block. `uvx` installs the two required workspace packages from
the public repository and caches the environment locally.
```json
{
"mcpServers": {
"grounded-cv": {
"command": "uvx",
"args": [
"--from",
"git+https://github.com/prathamesh-git9/mcp-servers.git#subdirectory=packages/grounded-cv",
"--with",
"git+https://github.com/prathamesh-git9/mcp-servers.git#subdirectory=packages/common",
"grounded-cv"
]
},
"repo-intel": {
"command": "uvx",
"args": [
"--from",
"git+https://github.com/prathamesh-git9/mcp-servers.git#subdirectory=packages/repo-intel",
"--with",
"git+https://github.com/prathamesh-git9/mcp-servers.git#subdirectory=packages/common",
"repo-intel"
]
},
"web-research": {
"command": "uvx",
"args": [
"--from",
"git+https://github.com/prathamesh-git9/mcp-servers.git#subdirectory=packages/web-research",
"--with",
"git+https://github.com/prathamesh-git9/mcp-servers.git#subdirectory=packages/common",
"web-research"
]
},
"ats-jobs": {
"command": "uvx",
"args": [
"--from",
"git+https://github.com/prathamesh-git9/mcp-servers.git#subdirectory=packages/ats-jobs",
"--with",
"git+https://github.com/prathamesh-git9/mcp-servers.git#subdirectory=packages/common",
"ats-jobs"
]
},
"outcome-ledger": {
"command": "uvx",
"args": [
"--from",
"git+https://github.com/prathamesh-git9/mcp-servers.git#subdirectory=packages/outcome-ledger",
"--with",
"git+https://github.com/prathamesh-git9/mcp-servers.git#subdirectory=packages/common",
"outcome-ledger"
]
},
"coding-workflows": {
"command": "uvx",
"args": [
"--from",
"git+https://github.com/prathamesh-git9/mcp-servers.git#subdirectory=packages/coding-workflows",
"--with",
"git+https://github.com/prathamesh-git9/mcp-servers.git#subdirectory=packages/common",
"coding-workflows"
]
}
}
}
```
## `.mcp.json`
```json
{
"mcpServers": {
"grounded-cv": {
"type": "stdio",
"command": "uvx",
"args": [
"--from",
"git+https://github.com/prathamesh-git9/mcp-servers.git#subdirectory=packages/grounded-cv",
"--with",
"git+https://github.com/prathamesh-git9/mcp-servers.git#subdirectory=packages/common",
"grounded-cv"
]
},
"repo-intel": {
"type": "stdio",
"command": "uvx",
"args": [
"--from",
"git+https://github.com/prathamesh-git9/mcp-servers.git#subdirectory=packages/repo-intel",
"--with",
"git+https://github.com/prathamesh-git9/mcp-servers.git#subdirectory=packages/common",
"repo-intel"
]
},
"web-research": {
"type": "stdio",
"command": "uvx",
"args": [
"--from",
"git+https://github.com/prathamesh-git9/mcp-servers.git#subdirectory=packages/web-research",
"--with",
"git+https://github.com/prathamesh-git9/mcp-servers.git#subdirectory=packages/common",
"web-research"
]
},
"ats-jobs": {
"type": "stdio",
"command": "uvx",
"args": [
"--from",
"git+https://github.com/prathamesh-git9/mcp-servers.git#subdirectory=packages/ats-jobs",
"--with",
"git+https://github.com/prathamesh-git9/mcp-servers.git#subdirectory=packages/common",
"ats-jobs"
]
},
"outcome-ledger": {
"type": "stdio",
"command": "uvx",
"args": [
"--from",
"git+https://github.com/prathamesh-git9/mcp-servers.git#subdirectory=packages/outcome-ledger",
"--with",
"git+https://github.com/prathamesh-git9/mcp-servers.git#subdirectory=packages/common",
"outcome-ledger"
]
},
"coding-workflows": {
"type": "stdio",
"command": "uvx",
"args": [
"--from",
"git+https://github.com/prathamesh-git9/mcp-servers.git#subdirectory=packages/coding-workflows",
"--with",
"git+https://github.com/prathamesh-git9/mcp-servers.git#subdirectory=packages/common",
"coding-workflows"
]
}
}
}
```
## Architecture
```mermaid
flowchart LR
C[Claude Desktop / MCP client] -->|JSON-RPC over stdio| G[grounded-cv]
C --> R[repo-intel]
C --> W[web-research]
C --> A[ats-jobs]
C --> O[outcome-ledger]
C --> CW[coding-workflows]
G --> H[BM25 + 384-D dense vectors + RRF]
G --> P[(Committed structured profile)]
R --> GH[Public GitHub REST API]
W --> RB[URL policy + robots.txt + extractor]
A --> ATS[Six public ATS APIs]
O --> SQ[(SQLite WAL / FULL sync)]
CW --> RC[Local repo structure + configured quality gates]
G & R & W & A & O & CW --> CORE["Shared typed failures, deadlines, redaction,<br/>rate limiting and cache"]
```
The repository is a seven-package workspace: one small shared core and one independently
installable package for each server. Every server exposes tools, resources, and at least
one prompt over the official MCP Python SDK's stdio transport.
## What a hiring manager can verify quickly
| Server | One-line proof of work |
|---|---|
| **grounded-cv** | Structured profile resources, genuine hybrid RAG, exact-span citations, and conservative claim verification. |
| **repo-intel** | Read-only repository, language, topic, README, CI, commit, issue, and code-search intelligence through public GitHub APIs. |
| **web-research** | Search plus main-content extraction with fail-closed robots checks, SSRF protection, redirect/size limits, and per-host pacing. |
| **ats-jobs** | One typed model across public Greenhouse, Lever, Ashby, Workable, SmartRecruiters, and Recruitee boards. |
| **outcome-ledger** | Durable deterministic idempotency keys and an honest `outcome_unknown` state after a simulated process crash. |
| **coding-workflows** | Unified-diff review, dependency-aware planning, bounded repo gates, traceback-to-file triage, and conventional commit synthesis. |
The complete machine-readable catalog is pinned at [`docs/manifest.json`](docs/manifest.json),
and the detailed protocol contract is in [`docs/API.md`](docs/API.md).
## Grounded CV: demonstrable hybrid RAG
The corpus uses **atomic evidence chunking**: one independently verifiable fact per
chunk, headings retained as metadata, no chunk crossing a section boundary. This makes
every returned span independently citable and prevents a partially relevant paragraph
from laundering an unsupported claim.
Two independent rankings run for every query:
1. Okapi BM25 over normalized terms and documented synonym expansion.
2. Cosine similarity over deterministic 384-dimensional dense feature-hash vectors
built from words and character trigrams.
Reciprocal rank fusion combines both lists with `RRF(k=60)`. The verifier then checks
claim-token coverage and numeric consistency against the retrieved spans. Supported
claims receive exact quotes; unsupported claims receive no citations.
Committed offline fixture, 10 queries, `k=5`:
| Metric | Result |
|---|---:|
| recall@5 | **1.0000** |
| MRR | **1.0000** |
| nDCG@5 | **0.9920** |
Reproduce it with `uv run grounded-cv-eval`. The fixture is committed beside the profile,
and CI executes the same harness assertions without downloading a model or calling a
network service.
## Trust and failure contract
- Every tool has generated input and structured output JSON Schemas backed by Pydantic v2.
- Every call has both an MCP-level deadline and an upstream HTTP timeout.
- Public upstream failures become typed values (`timeout`, `blocked`, `not_found`,
`rate_limited`, `upstream_error`, `parse_error`, and others); low-level exceptions never
cross a tool boundary.
- GitHub, web, and ATS text is explicitly marked `content_is_untrusted`; prompts instruct
clients not to execute instructions found inside it.
- Coding diffs, tracebacks, repository text, and gate output are untrusted. Quality gates
use shell-free argv execution, credential-reduced environments, and hard deadlines;
configured package-manager scripts should run only in repositories the caller trusts.
- URL credentials, private networks, authenticated scraping, mutation APIs, and
secret-shaped ledger fields are rejected. Optional `GITHUB_TOKEN` improves GitHub rate
limits but is never returned or logged.
- Tests run with outbound sockets disabled except loopback, require no keys, and use
protocol clients plus committed provider fixtures.
- The contract suite drives servers from `docs/manifest.json`, and a discovery test derives
the server list from `packages/` on disk and fails when the two disagree. A server added
without being declared would otherwise be skipped by every guard above while the suite
still reported success.
## Development
```bash
uv sync --all-packages --dev
uv run ruff check .
uv run ruff format --check .
uv run pytest
```
CI runs those three gates on Python 3.11, 3.12, and 3.13. Licensed under MIT.
TDQS
B3.1/5.0
Scored across 2 tools
Disambiguation5/5
The two tools have clearly distinct purposes: lookup retrieves evidence, while verify_claim checks claims against that evidence. There is no ambiguity or overlap in their roles.
Naming Consistency4/5
Both tool names are imperative verbs, but 'lookup' is a single word while 'verify_claim' follows a verb_noun pattern. This is a minor deviation, and the names are still predictable and readable.
Tool Count3/5
With only two tools, the set feels slightly thin for a production-oriented server. However, the narrow scope of evidence retrieval and verification may justify the small number, making it borderline appropriate.
Completeness4/5
The two tools form a logical pipeline: retrieve evidence first, then verify claims against it. Minor gaps might exist, such as a way to browse or filter evidence sources, but the core workflow appears covered.
Maintenance
ActivitySlowing
ResponsivenessNo issues