Installing Vedix for Gemini CLI
Run the agentic-research pipeline inside Gemini CLI. Skills + agents + MCPs
are bundled in a gemini-extension.json manifest — install via
gemini extensions install.
Prerequisites
- Gemini CLI ≥ 3.0
- Git
- Python 3.11+
uvx (pip install --user uv)- API access to
gemini-3.1-pro-preview and gemini-3-flash-preview (or
their stable 2.5 fallbacks)
1. Install via the Gemini extensions registry
gemini extensions install https://github.com/danilkotelnikov/vedix
This clones the repo to ~/.gemini/extensions/vedix/, reads
gemini-extension.json from the repo root, and registers the 9 MCP servers
2. Install the supporting Python package + literature MCPs
~/.gemini/extensions/vedix/plugins/vedix/scripts/install.sh
The script:
- pip-installs
mempalace and the core MCP requirements
- clones
JackKuo666/semanticscholar-MCP-Server to ~/.vedix/external/
- probes
uvx (for openalex/arxiv MCPs) and npx (for pubmed/annas/fetcher)
- runs
mempalace init ~/.vedix/palace
- runs the Vedix MCP self-test
3. Set required env vars
export OPENALEX_EMAIL="[email protected]"
export SEMANTIC_SCHOLAR_KEY="your-key-from-semanticscholar.org" # optional
export ANNAS_BASE_URL="annas-archive.gl" # optional
export ANNAS_DOWNLOAD_PATH="$HOME/Downloads/AA" # optional
export ANNAS_SECRET_KEY="your-key" # optional
Add to ~/.bashrc / ~/.zshrc / ~/.config/fish/config.fish for
persistence.
4. Restart Gemini CLI
gemini restart
Gemini picks up the new skill, MCPs, and the extension manifest.
5. Verify
> activate skill vedix
> use vedix to research linear regression on synthetic data
Caveat: no subagent support
Gemini CLI doesn't have Task/spawn_agent. The pipeline runs as a
single-session inline execution — each phase's agent prompt is followed
inline by the main session model.
Recommendation:
- Set Gemini's session model to
gemini-3.1-pro-preview for full pipelines
(heavy phases dominate).
- Use
gemini-3-flash-preview for partial intents (review-only, plot-only,
literature-only).
Per-phase model pinning is documented in agent frontmatter under gemini:
blocks for the day Gemini CLI gains subagent support; for now, the session
model handles all phases.
Per-role recommended models
| Role group | Gemini model | Thinking | Output cap | Context |
|---|
| Heavy (ideator, hypothesizer, code-generator, manuscript-writer, reviewer) | gemini-3.1-pro-preview | thinking_level: high | 65,536 | 2,000,000 |
| Light (codebase-scanner, literature-searcher, experiment-runner, plotter, citator, meta-analyst) | gemini-3-flash-preview | thinking_budget: 8192 | 8,192 | 1,000,000 |
| Fixer | gemini-3-flash-preview | thinking_budget: 16384 | 16,384 | 1,000,000 |
Updating
gemini extensions update vedix
Uninstalling
gemini extensions uninstall vedix