Continuous Claude v3: A Persistent, Multi‑Agent Open‑Source Dev Environment
What Is Continuous Claude v3?
Continuous Claude v3, hosted on GitHub as parcadei/Continuous‑Claude‑v3, is a full‑stack framework that extends Claude Code into a persistent, learning development environment. It solves three core pain points in current LLM‑powered coding assistants:
- Context loss – When Claude’s prompt window fills, the agent compacts the conversation, wiping out nuances. Continuous Claude keeps a continuity ledger in a PostgreSQL database so state survives across sessions.
- Token waste – Reading entire files or sending raw source to an LLM is expensive. The built‑in TLDR code analysis (five‑layer AST, CFG, DFG, PDG, etc.) returns a compressed 1,200‑token summary instead of 23,000 tokens.
- Scattered tooling – One tool, a dozen plugins, a million hooks. Continuous Claude bundles 109 skills, 32 agents, and 30 lifecycle hooks that work together to orchestrate multi‑step workflows without the user needing to remember slash commands.
In short, Continuous Claude turns Claude Code from a “chat‑based” assistant into a continually learning, multi‑agent workbench.
Core Concepts & Architecture
| Component | Purpose | Key Details |
|---|---|---|
| Skills | Individual actions (e.g., fix, build, pro) |
109 modular plugins, each defined in ~/.claude/skills/ |
| Meta‑Skills | Orchestrated workflows | /fix, /build, /tdd, etc. are chains of agents & tools |
| Agents | Specialized sub‑processes | 32 agents like kraken, schleud, oracle, each in ~/.claude/agents/ |
| Hooks | Lifecycle triggers | 30 hooks, e.g., session-start, pre‑tool‑use, post‑tool‑use |
| TLDR Code Analysis | Token‑efficient source representation | 5‑layer analysis (AST → PDG) + semantic index |
| Memory System | Cross‑session learning | PostgreSQL + pgvector stores sessn, archival_memory, handoffs |
| Continuity Ledger | In‑session state | Markdown ledger located in thoughts/ledgers/ |
| Handoffs | Transfer between sessions | YAML files in thoughts/shared/handoffs/ |
The entire stack runs in Docker (PostgreSQL) with a Python/TypeScript codebase. The setup wizard (scripts/setup.wizard) configures environment vars, builds the database, installs the 109 skills, and optionally installs SymPy, Z3, Pint, and Lean4 for math/verification.
How to Install & Set Up
Below is a step‑by‑step guide that mirrors the official wizard, ideal for developers who prefer manual control.
# 1. Clone the repo (one‑time)
git clone https://github.com/parcadei/Continuous-Claude-v3.git
cd Continuous-Claude-v3/opc
# 2. (Optional) Create a virtual environment
uv venv .venv
source .venv/bin/activate
# 3. Run the interactive setup wizard
echo "Running Continuous Claude setup wizard..."
uv run python -m scripts.setup.wizard
The wizard walks you through:
1. Backing up any existing .claude/ config.
2. Docker checks – installs PostgreSQL container.
3. Database URL – sets CONTINUOUS_CLAUDE_DB_URL.
4. Skill & hook installation – copies 109 skills & 30 hooks.
5. Optional features – installs math libraries, TLDR, Loogle, etc.
6. PostgreSQL init – creates the four core tables (sessions, file_claims, archival_memory, handoffs).
Once finished, start the tool by running:
claude # or, if using the full CLI wrapper
If you ever need to update the repository or its hooks/skills, simply:
cd Continuous-Claude-v3/opc
uv run python -m scripts.setup.update
First‑Time Workflow Example
After installation, launch Claude Code and run a basic workflow to see the magic.
> /fix bug "Login route throws 500 on bad creds"
Workflow steps:
1. sleuth – searches logs, source, and uses ast-grep.
2. premortem – identifies risk factors.
3. kraken – runs tests in sandboxed environment.
4. test – verifies fixes.
5. commit – auto‑generates a pull request with diff and commit message.
You’ll notice near‑zero token usage because the TLDR hook injects summarized code snippets rather than raw files.
Why Continuous Claude Is a Game‑Changer
| Benefit | Impact |
|---|---|
| Persistent State | Work on a feature for hours, close the terminal, and pick up exactly where you left off – thanks to continuity ledgers and handoffs. |
| Token Efficiency | TLDR reduces ~23k raw tokens to ~1.2k, slashing prompt costs and speeding up responses. |
| Zero Cost | The entire stack is MIT‑licensed, open‑source, and relies only on free services (PostgreSQL, Docker). Optional APIs (Perplexity, NIA) are non‑mandatory. |
| Built‑in CI/CD | Hooks run ruff and pyright automatically on edits, providing instant linting before you commit. |
| Extensible Workflows | Add or modify custom skills and agents in seconds—no need to re‑train models. |
| Community‑driven | With 3.2k stars and an active contributors list, the ecosystem continues to grow. |
Contributing & Extending
Contributions are encouraged! The project is structured so you can drop new skills into ~/.claude/skills/, add agents, or improve hooks in ~/.claude/hooks/. Remember to run the update script to sync your changes.
Key entry points:
- skills – skill-rules.json for activation triggers.
- hooks – TypeScript (src/) compiled to dist/. Use the tldr hook to extend token‑efficiency.
- agents – Extend with new orchestrator patterns in agents/.
Read the CONTRIBUTING.md for guidelines, and fork the repo to experiment in isolation.
Final Thoughts
Continuous Claude v3 bridges the gap between powerful LLM code assistants and real‑world developer workflows. With persistence, token‑efficiency, and a modular architecture, it frees you to focus on building rather than juggling context or hunting down bugs across multiple files.
If you’re using Claude Code and have felt the context‑compaction pain, or you’re simply looking for a feature‑rich, community‑supported AI dev platform, Continuous Claude v3 is ready to jump into your workflow tomorrow. Clone, run the wizard, and let the agent ecosystem take care of the heavy lifting.
Happy hacking!}