Documentation

Install Senteron as a Claude Desktop extension or as a CLI tool. Both ship from the same repository.

Repository setup

The MCP extension and the CLI both delegate to the same script, so step one is to get the repo working on its own.

git clone https://github.com/senteron/senteron.git
cd senteron
python3 -m venv .venv
source .venv/bin/activate          # Windows: .venv\Scripts\activate
pip install -r requirements.txt
cp .env.example .env               # then edit .env and add your API keys
python3 senteron.py --no-optimize "ping test"

If the last line returns a synthesis, the CLI is good to go and you can use it directly or move on to the MCP extension.

Claude Desktop extension

The recommended way to use Senteron from Claude Desktop is the pre-built bundle:

1. Download senteron-0.1.0.mcpb from the repo's dist/ directory.
2. Open Claude Desktop → Settings → Extensions.
3. Drag the .mcpb file into the window (or click Install from file).
4. When prompted for Senteron repository path, enter the absolute path to your local repo from the previous section (the directory containing senteron.py, .venv/, and runs/).
5. Optional: paste your provider API keys (ANTHROPIC_API_KEY, OPENAI_API_KEY, GEMINI_API_KEY, and seven more) into the configuration fields. Keys you fill in here take precedence over the same keys in your repo's .env file. Fields you leave blank fall through to .env as before. One caveat: keys entered here only reach the Claude Desktop extension — running senteron.py from a terminal, or the web UI (./start-ui.sh), still reads from .env. If you use multiple surfaces, fill .env too.
6. Enable the extension. Claude Desktop's bundled uv installs the MCP SDK on first run; no pip install needed for the bundle itself.

Four MCP tools

Smoke test sequence

After installing the extension, run these four prompts in a fresh Claude Desktop conversation. Each step verifies one capability.

1. Sanity check. Run senteron_health and show me the full result. Expect ok: true, all subsystems passing, and a non-zero runs_count.
2. Read existing runs. List my last 5 Senteron runs. Expect five rows, each with readable: true.
3. Inspect one run. Pass a run id from step 2 to senteron_get_run.
4. End-to-end run. Run Senteron on this prompt: "What are the key trade-offs between REST and GraphQL?" Use profile=quick, pipeline=judge_only, no_optimize=true. Expect 1–2 minutes, a populated synthesis, and a new file under runs/.

Long-running peer-review runs

The full peer-review pipeline (profile=standard and above) takes 5–10 minutes; profile=max can run 15+. Senteron's MCP server streams info + report_progress notifications for each stage transition so Claude Desktop's per-tool-call deadline resets continuously and long runs do not get dropped.

For runs that exceed even the streaming-extended budget, the CLI is the escape hatch:

.venv/bin/python3 senteron.py --profile opus --no-optimize \
  --timeout 180 "your prompt here"

Then ask Claude via MCP to fetch the newest saved run with senteron_list_runs + senteron_get_run.

Manual configuration (no bundle)

Use this only for development or for MCP clients without bundle support. Add to claude_desktop_config.json:

{
  "mcpServers": {
    "senteron": {
      "command": "/absolute/path/to/senteron/.venv/bin/python3",
      "args": ["-m", "mcp_server"],
      "cwd": "/absolute/path/to/senteron"
    }
  }
}

You'll also need pip install -r requirements-mcp.txt in the repo venv.

Source & issues

Full source, the integration plan, and the issue tracker live at github.com/senteron/senteron.