Make your AI agent smarter at startup investing

SICTIC-AI works inside the AI agent you already use, such as Codex, Claude Code or Cowork, OpenClaw, or Gemini. It gives your agent reusable skills for startup analysis, due diligence, investor matching, and angel-network operations. The important checks and workflows are codified up front, so your agent can apply them consistently instead of inventing an approach on the spot.

To set it up, ask your agent to install this repository. It can inspect your machine, install the required local services, configure the toolkit, and verify the result. A ready-to-use installation prompt is provided below.

Once installed, run a skill by asking naturally: “Run startup_profile for SpaceX” or “Ask the spacex dataset about its main technical risks.” Your agent reads the relevant skill instructions, runs the workflow, and returns the result with its supporting evidence.

Built from real angel-investing workflows

SICTIC is Switzerland's largest and most active angel investor network. It brings together more than 500 investors, has hosted more than 1,000 startup pitches, and its investor community has helped fund more than 300 Swiss technology startups (sictic.ch). SICTIC-AI turns that practical experience into an open-source toolkit that is free to use and contribute to.

It helps:

• Startups: Review funding materials before approaching investors.
• Business angels: Accelerate common due-diligence assessments.
• Angel investor networks: Support member engagement, startup selection, due diligence, and portfolio monitoring.
• Contributors: Improve how AI supports early-stage funding by extending shared, reviewable workflows.

Recent releases also add hybrid semantic and keyword retrieval, table-aware spreadsheet ingestion, safer Qdrant lifecycle management, and dependency-aware bulk refreshes.

Install with your AI agent

Start your agent and send it this prompt:

Install SICTIC-AI for me using its default local setup.

Before running commands, ask me exactly one question:
“Which exact local folder should contain the SICTIC-AI repository?
For example: /Users/you/SICTIC-AI”

After the initial folder question, ask me again only if:
- the selected folder cannot be used safely;
- installing a system prerequisite requires my authorization;
- the operating system cannot support the default setup; or
- setup fails and there are multiple materially different ways to proceed.

Do not configure Google Drive synchronization, hosted model providers, API
keys, or other optional integrations. Mention them only after the local
installation succeeds.

After I answer, continue autonomously:

1. Check whether the operating system is macOS, Linux, or WSL2.
2. Check the tools needed to clone the repository:
   - on macOS, first check whether Homebrew is installed, then check Git;
     install either one if missing
   - on Ubuntu, Debian, or WSL2, check Git, curl, and wget; install any that are
     missing using apt
3. Clone https://github.com/ducroo/SICTIC-AI into the folder I selected. If
   that folder already contains the correct repository, reuse it after
   verifying its Git remote. Never overwrite an unrelated or modified folder.
4. Read README.md and follow its “Manual setup” section in sequence.
5. Check for Miniforge/Conda and Ollama. Install either one if missing using the
   documented platform commands.
6. Run the installer without copying skills to an agent skills directory, using
   these defaults:
   - REPO_PATH: the selected repository folder
   - LOCAL_STORAGE_PATH: <REPO_PATH>/local_storage
   - LOCAL_DATA_PATH: <REPO_PATH>
   - accept the installer's default local models, URLs, and service settings
7. Start Ollama and Qdrant with ./launch.sh start and confirm that both services
   are running. Allow the launcher to pull configured Ollama models that are
   missing, but do not delete or reinstall models that are already present.
8. Run the test suite and the skill harness help command to verify the setup.
9. Report what was installed, the important paths, and one example command for
   running a user-facing skill.

Do not modify tracked repository source files as part of installation.

This prompt deliberately installs local models, so it does not require an API account or send startup data to a hosted LLM provider.

Switching models

The bundled local models make the default installation widely accessible, but more capable models are available if your hardware or budget permits. Local options include the Gemma 4 and Qwen 3.6 families through Ollama; cloud options include gpt-5.6-luna through the OpenAI API. Larger local models generally require substantially more memory and disk space.

The example below switches text generation to OpenAI while retaining the local vision and embedding models. Other local models and cloud providers follow a similar route: select the model and configure its endpoint and API key when required.

To use gpt-5.6-luna:

1. Create an account at platform.openai.com, configure billing, and create an API key.
2. Replace these values in .env:

LLM_MODEL=openai/gpt-5.6-luna
LLM_BASE_URL=
LLM_API_KEY=your-api-key

Treat the API key like a password: never commit or share your .env file. The model name and API-key workflow are documented in the official OpenAI model documentation and API quickstart.

Contributing

Read AGENTS.md for working procedures and the standards skill for technical contracts. Each skill's SKILL.md describes its own workflow; prompts and assessment criteria live in config/.

By establishing shared standards for early-stage investing, we aim to structurally strengthen the startup ecosystem in Switzerland and Europe.

Join us with your experience, questions, and ideas! Critical thinking and practical investing experience are the differentiators; We already have the AI wizards in the team.

We develop skills in teams and meet regularly to challenge assumptions, compare results, and improve the workflows. Help us turn strong investing practice into open, reusable skills.

Manual setup

SICTIC-AI requires macOS, Linux, or WSL2, plus Git, Miniforge, and Ollama. Ollama is required for the default local models. Install the prerequisites for your platform first.

On macOS:

# Install Homebrew first if `brew` is not available:
/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"

brew install --cask miniforge
conda init zsh
exec $SHELL
brew install ollama

On Ubuntu, Debian, or WSL2:

sudo apt update && sudo apt install -y git curl wget
wget "https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-$(uname)-$(uname -m).sh"
bash Miniforge3-*.sh
conda init bash
exec $SHELL
curl -fsSL https://ollama.com/install.sh | sh

Then clone the repository and run the installer:

git clone https://github.com/ducroo/SICTIC-AI.git
cd SICTIC-AI
./install.sh
./launch.sh start
./launch.sh status

The installer creates or updates the sictic-env Conda environment and helps configure .env. It also asks for INSTALLED_SKILLS_PATH: the optional directory into which it copies skills for discovery by your AI agent.

• Choose none to work directly from the repository without copying skills.
• Enter an absolute path to make SICTIC-AI skills discoverable by that agent.

For the default local setup:

• Set REPO_PATH to /Users/you/SICTIC-AI, replacing it with the folder where you put the repository.
• Accept the suggested local storage and data paths.
• Choose none for INSTALLED_SKILLS_PATH.
• Accept the local Ollama model, URL, and service defaults.

The installer writes these choices to .env.

Repository-only mode still supports all commands documented below. Re-run the installer after editing or adding skill instructions if you use copied skills.

The commands above are sufficient for the default local setup. For hosted model providers, advanced command interfaces, and maintenance tasks, see Installation and operations.

Configuration paths

The installer suggests sensible values for a new .env:

Variable	What it controls	Example
REPO_PATH	Root of this Git repository and its source code	/Users/you/SICTIC-AI
INSTALLED_SKILLS_PATH	Optional directory receiving agent-discoverable copies of skills	none or /Users/you/.claude/skills
LOCAL_STORAGE_PATH	Local root containing startup, community, and generated datasets and insights	/Users/you/SICTIC-AI/local_storage
LOCAL_DATA_PATH	Machine-local root under which cache/ and docling_data/ are stored	/Users/you/SICTIC-AI

All skills operate on LOCAL_STORAGE_PATH. Optional Google Drive synchronization is independent of the toolkit runtime and is not configured by install.sh. Model and service variables are explained in .env-template and in the operations guide.

Running a skill

Ask your AI agent, or use the command harness directly:

conda run -n sictic-env python -m skills.harness /startup_profile SpaceX

conda run -n sictic-env python -m skills.harness /dataset_chat SpaceX "What are the main risks?"

conda run -n sictic-env --no-capture-output python -m skills.harness

Run /help in the interactive harness to see available commands. Some administrative skills use their own python -m skills.<name> interface; their SKILL.md files and the operations guide show the supported syntax.

User-facing skills

✅ means available. 🚧 means work in progress. ◻️ means planned but not started. Internal building blocks are intentionally omitted.

Skill	Status	Description
Community
expert_search	✅	Finds members with relevant domain expertise for due diligence or operational support.
potential_investors	✅	Finds investors with the strongest fit for a startup.
advocates	✅	Finds members suited to representing the organization at external events.
investor_profile	✅	Combines a member's professional profile, investment record, and preferences.
suggested_startups	✅	Ranks selected stored startup profiles for each investor; default dataset selection does not verify fundraising status.
Startup selection and jury
submission_ready	🚧	Checks whether a Dealum application is complete and meets initial eligibility criteria.
pitch_ready	🚧	Assesses whether a startup is mature enough and ready to pitch at a SICTIC event.
Due diligence
dataset_chat	✅	Answers evidence-based questions about a startup or community dataset.
startup_profile	✅	Produces a concise, neutral overview used by other skills.
team_profile	✅	Assesses founders and the overall team.
person_profile	✅	Creates a comprehensive profile of a founder, member, or other person.
startup_traction	✅	Summarizes and quantifies commercial traction.
dd_checks	✅	Runs a broad suite of due-diligence checks.
dd_priorities	✅	Synthesizes up to eight decision-relevant priorities from a saved dd_checks report.
startup_website_import	✅	Imports a startup's public website into its due-diligence dataset.
market_review	◻️	Reviews market size, customer needs, competition, and substitutes.
sha_review	🚧	Reviews a selected Shareholders' Agreement against a reference SHA and legal checklists.
companyresearch.ch	🚧	Uses the companyresearch.ch API to collect publicly available information about a startup.
Ongoing monitoring
alerts_and_news	◻️	Monitors and interprets relevant portfolio-company news and updates.
startup_support	◻️	Coordinates operational support from investors.
portfolio_mgmt	◻️	Produces portfolio risk, return, and performance overviews.
Data and operations
dealum_import	✅	Imports the application dossier of a startup from the Dealum.com platform using an API key.
bulk_refresh	✅	Refreshes selected insights across selected datasets.
dataset_maintenance	✅	Diagnoses, migrates, prunes, and repairs datasets and search indexes.
linkedin_maintenance	✅	Finds missing LinkedIn profiles and imports manually collected profiles.

Where is my data?

Application data is rooted at LOCAL_STORAGE_PATH; durable parsed documents and disposable runtime data are rooted at LOCAL_DATA_PATH.

Path	Description
<LOCAL_STORAGE_PATH>/storage/startups/<startup_name>/	The collection of information about a single startup.
↳ ./datasets/	Startup data room: pitch decks, spreadsheets, PDFs, website imports, and other source material.
↳ ./insights/	Generated startup reports.
<LOCAL_STORAGE_PATH>/storage/community/<community_name>/	A community or member collection, such as sictic-members.
↳ ./datasets/	Community and member source data.
↳ ./insights/	Generated community reports and profiles.
<LOCAL_STORAGE_PATH>/storage/generated/<dataset_name>/	A searchable dataset assembled from generated insights.
↳ ./datasets/	Materialized source documents for the generated dataset.
↳ ./insights/	Reports associated with the generated dataset.
<LOCAL_DATA_PATH>/	Machine-local parsed data and disposable runtime data.
↳ ./docling_data/	Durable parsed documents; not synchronized to cloud storage.
↳ ./cache/	Disposable runtime cache and temporary state.
↳ ./cache/scheduler.json	Shared concurrency state for model and Docling jobs.

For example, place a pitch deck in <LOCAL_STORAGE_PATH>/storage/startups/spacex/datasets/. Running startup_profile writes its result below <LOCAL_STORAGE_PATH>/storage/startups/spacex/insights/.

Where to learn more

The docs/ folder currently contains:

• Installation and operations: detailed installer modes, environment and model configuration, background services, command interfaces, retrieval, maintenance, and tests.
• Codebase assessment: a historical architecture review; use current standards and skill documents for implementation contracts.

Optional Google Drive synchronization

The toolkit itself always reads and writes local files. If you want to share the application-storage tree through Google Drive, the optional rclone-sync helper provides guarded bidirectional synchronization while converting local Markdown files to native Google Docs and exporting Google Docs back to Markdown.

Cloud access is deliberately user-owned. Install rclone, create and authenticate your Google Drive remote with rclone config, and then run the guided repository setup:

./rclone-sync/configure.sh
./rclone-sync/rclone-sync.sh bootstrap-dry-run

Review the dry-run output carefully before establishing the first baseline:

./rclone-sync/rclone-sync.sh bootstrap

After bootstrap, preview and run routine synchronization with dry-run and sync. See rclone synchronization for installation, recovery, safety, and scheduling details. Existing legacy cloud variables in .env are ignored and may be removed manually after the rclone setup has been verified.