ChemSpace Copilot

Warning This repository is under active development. APIs, agent behavior, and project structure may change without notice.

---

Overview

ChemSpace Copilot combines language-model coordination with deterministic cheminformatics tools. The default Agno team coordinates seven specialists for ChEMBL retrieval, GTM workflows, downstream analysis, reporting, molecular design, peptide design, and retrosynthesis. An optional MCP server exposes the same scientific core to external reasoning clients. A separate robustness agent and a single-agent ablation baseline are registered outside the production team. The GTM engine is provided by ChemographyKit.

Chainlit / CLI ──► Agno coordinator ──► 7 specialists ──┐
                                                       ├─► tools ─► events/artifacts
External client ──► MCP profiles and adapters ─────────┘
                         ▲                    ▲
                         └── skills/workflows ┘

Features

• 7 Runtime Agents + Evaluation/Ablation Factories — Seven production specialists, a separate robustness evaluator, and a controlled single-agent baseline
• Generative Topographic Mapping — Dimensionality reduction and visualization of chemical space via ChemographyKit
• Molecular and Peptide Generation — Molecular Designer small-molecule generation with autoencoder and LLM engines plus Peptide Designer generation with WAE and LLM engines, interpolation, and GTM-guided targeting
• S3/MinIO Integration — Session-scoped cloud storage with local filesystem fallback
• Chainlit Interface — WebSocket-based real-time chat with password authentication, file upload, and inline molecule rendering
• Artifact-backed Runs — Event-sourced workflow state, checksummed artifacts, and session-local chat context; cross-session agent memory is disabled
• Robustness Testing — Framework for validating prompt variation handling with semantic similarity scoring

Quick Start

Option 1: Docker (Recommended)

# Build containers
docker compose build

# Run (prompts for DEEPSEEK_API_KEY only when using the DeepSeek provider)
./docker-start.sh

Access the application at http://localhost:8000

To remove the docker-compose installation and rebuild it from scratch use:

docker compose down --volumes --remove-orphans --rmi all

WARNING! This would wipe all the previously generated assets on S3 and the previously run sessions of the chatbot.

See the Docker guide for the full Docker deployment guide.

Option 2: Local Installation

For file-based configuration, copy .env.example to .env in the project root:

Before the first local Chainlit run, generate a Chainlit auth secret and copy the generated key into .env as CHAINLIT_AUTH_SECRET:

uv run chainlit create-secret

CHAINLIT_AUTH_SECRET=your-generated-key-here

# Required only for the default DeepSeek provider
DEEPSEEK_API_KEY=your-api-key-here

# Optional model overrides (otherwise .modelconf is used)
# MODEL_PROVIDER=deepseek
# MODEL_ID=deepseek-chat
# OLLAMA_HOST=http://localhost:11434

# Optional — S3/MinIO storage (set USE_S3=true only when you want remote storage)
USE_S3=false
# When enabled:
S3_ENDPOINT_URL=http://localhost:9000
MINIO_ACCESS_KEY=cs_copilot
MINIO_SECRET_KEY=chempwd123
ASSETS_BUCKET=chatbot-assets

# Optional — ChEMBL local MySQL (faster queries, offline use)
# Download dump: https://chembl.gitbook.io/chembl-interface-documentation/downloads
# CHEMBL_MYSQL_HOST=localhost
# CHEMBL_MYSQL_PORT=3306
# CHEMBL_MYSQL_USER=chembl
# CHEMBL_MYSQL_PASSWORD=
# CHEMBL_MYSQL_DATABASE=chembl_36

The repository also includes a tracked .modelconf file. Edit it if you want to switch from the default DeepSeek backend to a local Ollama model.

uv sync --frozen

Retrosynthesis uses the optional SynPlanner 1.7.0 profile:

uv sync --frozen --extra retrosynthesis

Its published dependencies support Linux x86_64 and macOS ARM64, but currently lack Linux ARM64 wheels. The core application runs without this extra. When running with extras, use uv run --no-sync to retain the installed profile. See the installation guide for details.

# Run the interactive setup script
python scripts/setup_s3.py

# Or build and start the pinned MinIO source image manually
docker build -f Dockerfile.minio --target server -t cs_copilot-minio:local .
docker run -d --name minio \
  -p 9000:9000 -p 9001:9001 \
  -v /mnt/data:/data \
  -e MINIO_ROOT_USER=cs_copilot \
  -e MINIO_ROOT_PASSWORD=chempwd123 \
  cs_copilot-minio:local server /data --console-address ":9001"

If the container already exists: docker start minio

Chainlit persistence is disabled by default in chainlit.toml. Only set up PostgreSQL if you plan to enable Chainlit persistence manually.

docker run --name chainlit-pg -p 5432:5432 -d \
  -e POSTGRES_PASSWORD=postgres \
  -e POSTGRES_USER=postgres \
  -e POSTGRES_DB=chainlit \
  postgres:16

export DATABASE_URL="postgresql://postgres:postgres@localhost:5432/chainlit"

If the container already exists: docker start chainlit-pg

Usage

Chainlit App

uv run chainlit run chainlit_app.py -w

Notes:

• The bundled chainlit.toml currently has [persistence] enabled = false.
• The default local login credentials are username admin and password admin123 (defined in chainlit_app.py). Change these before exposing the app beyond local development.
• The app sets a per-thread title from your first message; you can rename it in the UI.

Jupyter Notebook

An example workflow is available in notebooks/cs_copilot.ipynb.

MCP server (optional)

An optional Model Context Protocol server lets external MCP clients (Codex, Claude Code) drive cs_copilot toolkits, prompts, and session artifacts directly — the external client is the reasoning engine; the Agno multi-agent team is not invoked. The default Chainlit and CLI runtimes are unaffected.

uv sync --extra mcp

# Preflight the remote HTTP path ChatGPT/tunnels will call
cscopilot-mcp-check

# Local stdio MCP clients (Codex, Claude Code)
cscopilot-mcp --profile gtm-analysis --session-id demo --workflow-slug gtm-density-landscape

# Remote MCP clients (ChatGPT apps, browser-hosted clients)
cscopilot-mcp-serve --profile gtm-analysis --session-id demo --workflow-slug gtm-density-landscape --host 127.0.0.1 --port 8000
# Add --allowed-host <your-host> when a reverse proxy preserves the public Host header.

# Optional bearer-token protection for HTTP clients/proxies that send Authorization
CS_COPILOT_MCP_AUTH_TOKEN=change-me cscopilot-mcp-serve --profile standard \
  --session-id demo --workflow-slug chembl-to-gtm-report \
  --host 127.0.0.1 --port 8000 --allowed-host <your-host>

For ChatGPT, expose the streamable HTTP endpoint (/mcp) through a reachable HTTPS URL or an approved secure tunnel. If a proxy forwards public Host or Origin headers, pass them with --allowed-host / --allowed-origin. The server includes read-only search / fetch tools for ChatGPT data-only/deep-research compatibility, plus the full cs_copilot tool catalog for full MCP developer-mode clients. Tool descriptors include MCP readOnlyHint annotations so ChatGPT can distinguish pure lookup tools from state-changing cs_copilot workflows; network-capable tools also advertise openWorldHint for approval policy. The selected startup profile is a strict discovery/invocation boundary; narrower profiles include chembl-retrieval, gtm-analysis, chemoinformatics, reporting, molecular-design, peptide-design, retrosynthesis, and robustness. By default, LLM-dependent MCP tool paths create llm_* tasks for the external client to complete. Trusted private deployments can pass --llm-policy agno-model to load only the configured Agno model for toolkit LLM calls, without enabling the Agno team.

Example client configs ship under examples/mcp/ (claude_code.json, codex.toml, codex_http.toml, chatgpt_remote.md, codex_subscription_smoke.md, secure_mcp_tunnel.md). See docs/mcp.md for the full tool / prompt / resource catalog and the ChEMBL LLM-as-judge gating contract.

The repository also contains a Codex plugin under plugins/chemspace-copilot/. It packages the standard MCP profile and a thin bootstrap skill; scientific procedures remain in skills/ and workflow_catalog/, not in the plugin. This is intentionally a repo-local marketplace delivery surface (.agents/plugins/marketplace.json), not Python wheel content. Before installing it in Codex, install the MCP extra (uv sync --extra mcp, or pip install "cs_copilot[mcp]") and ensure cscopilot-mcp is on the environment's PATH. Then open that marketplace file in Codex using its absolute checkout path and install chemspace-copilot; a portable deep link cannot be committed because every clone has a different local path.

Architecture

The agent registry contains nine factories: seven production team members, a separate robustness evaluator, and a single-agent architecture-ablation baseline. The Chainlit/CLI path uses the Agno coordinator as its reasoning engine; in MCP mode, the external client supervises capability-filtered tools.

Runtime Team

Agent	Role
ChEMBL Downloader	Downloads and filters bioactivity data from ChEMBL (REST API by default; optional local MySQL backend)
GTM Agent	Unified GTM operations: build, load, density analysis, activity landscapes, projection, and GTM sampling support
Chemoinformatician	Downstream chemoinformatics analysis including scaffold, similarity, clustering, and SAR workflows
Report Generator	Formats analysis results into reports and visual outputs
Molecular Designer	Small-molecule design via autoencoder and LLM engines, including standalone and GTM-guided modes
Peptide Designer	Peptide design via WAE and LLM engines, latent-space GTM workflows, and DBAASP-backed peptide activity landscapes
SynPlanner	Retrosynthetic planning and route visualization for target molecules

Separate Evaluation Agent

Agent	Role
Robustness Evaluation	Analyzes robustness test runs, score distributions, failures, and trends

Single-Agent Baseline

The single_agent factory combines the production tool surface into one flat agent for controlled multi-agent-versus-single-agent robustness experiments. It is not part of the production team.

Agents share within-session state via session_state and can persist bounded conversation history in SQLite. Cross-session user and agentic memories are disabled. The default Chainlit and CLI team uses structured but process-local ad-hoc handoffs. When a v2 RunContext is supplied—or an MCP client drives the v2 lifecycle—scientific run state is stored as replayable events and checksummed artifacts through the unified S3/local storage abstraction. Strict persisted task-DAG enforcement is currently the chembl-to-gtm-report MCP pilot; other catalog workflows remain taskless/legacy.

For full architectural details, see the Architecture Overview and Skills and Workflows.

License

This project is licensed under the MIT License.

Contributing

Contributions are welcome! Please:

1. Fork the repository
2. Create a feature branch (git checkout -b feature/my-feature)
3. Ensure code passes pre-commit run --all-files
4. Submit a pull request

See the Contributing Guide for code style conventions and detailed guidelines.