OpenConclave

Where one agent isn't enough.

Run multiple AI agents together — in pipelines, debates, and teams. They work in parallel, challenge each other, ask you questions mid-run, and remember what they learned. All on your machine.

---

Install

As a Claude Code plugin (recommended):

/plugin marketplace add openconclave/oc
/plugin install openconclave@openconclave

Starts the OC server automatically, wires up MCP tools for managing conclaves, and delivers conclave events to your Claude Code session as notifications. Run /openconclave:open (or visit localhost:4000) to open the editor.

Requires Bun on your PATH. See Claude Code plugin below for details.

Standalone binary (no Claude Code required):

# Windows
irm https://openconclave.com/install.ps1 | iex

# macOS / Linux
curl -fsSL https://openconclave.com/install.sh | bash

Open localhost:4000 → import a starter → hit Run. No runtime dependencies. Single binary. Manual install →

---

Try something

Browse the conclaves repo for ready-to-import starters — deep code reviews, multi-agent debates, decision-support tools, and more. Drop a .json file into your OC instance and run it.

To import: Conclaves → ⬇ Import → paste URL or drop the .json file → map provider roles → Run.

---

Two ways to use it

Pipelines

Chain agents into workflows with a visual editor. Agents run in parallel, merge results, branch on conditions, loop until tests pass, and pause to ask you questions.

Trigger → Agent → Output                          (simple)
Trigger → [Agent, Agent, Agent] → Merge → Agent   (parallel specialists)
Trigger → Agent → Condition → Agent ↺              (retry loop)
              ↕
         Channel Loop                               (ask Claude Code)

Use cases: code review, TDD pipelines, security audits, data validation, content generation with editorial review — anything where one perspective isn't enough and you want the system to learn over time.

Discussions

Put agents in a room with a moderator. They debate, challenge each other, and build on ideas across multiple rounds. The moderator steers, summarizes, and ends with a verdict.

Use cases: brainstorming, red-teaming, scenario planning, decision analysis, exploring a topic from multiple angles — or just entertainment.

---

What makes this different

Agents learn across runs. A knowledge base stores lessons from every run. Next time, agents consult those lessons before acting. Your pipelines get smarter the more you use them.

Multiple models in one workflow. Use Claude for deep reasoning, GPT for a second opinion, Ollama for local-only steps — in the same pipeline. Each agent picks its own model.

Claude-in-the-loop. When an agent hits an ambiguous decision, it pauses the pipeline and asks Claude Code via the channel loop. Claude Code answers from context — or escalates to you in the terminal. Not a checkbox approval — an intelligent intermediary that only bothers you when it actually needs to.

Everything stays local. Single binary, runs on your machine, your code never leaves your laptop. No cloud. No account.

Pipelines become tools. Every conclave exposes as an MCP tool. Claude Code can trigger your pipelines, receive results, and answer channel loop questions.

---

Build your own

Drag nodes, draw connections, write system prompts.

Node types: Trigger · Agent · Condition · Code · Merge · Channel Loop · Output · File · Discussion · Knowledge

AI engines: Claude · OpenAI · Ollama · any OpenAI-compatible API

Agent tools: Bash · Read · Write · Edit · Glob · Grep · WebSearch · WebFetch · Knowledge Search · MCP tools

Triggers: Manual · Cron · Webhook · Telegram · Claude Code channel

Knowledge bases: Embed documents with Ollama, agents get search_knowledge as a tool. Agents can write lessons back — your pipeline improves with every run.

Git isolation: Pipelines that modify code run in worktrees. Your main branch is never touched.

---

Requirements

• Ollama — for local models and embeddings: ollama pull nomic-embed-text
• API keys (optional) — ANTHROPIC_API_KEY, OPENAI_API_KEY, or any OpenAI-compatible provider
• Bun — required when installing via the Claude Code plugin (the plugin runs OC from source). The standalone binary install doesn't need it.
• Claude Code (optional) — for the plugin workflow described below.

Claude Code plugin

The plugin is a single-package install that bundles the OC server, the editor UI, and the MCP surface for managing conclaves from Claude. It replaces the older pair of plugins (openconclave-channel + openconclave-dev).

What it does

• Starts the server automatically via a Claude Code background monitor. Your editor is live at http://localhost:4000 for the duration of the session.
• Exposes an MCP server named openconclave with tools to list/create/update conclaves, trigger runs, manage the scheduler, and respond to blocked runs.
• Delivers conclave events as notifications. When a conclave emits output or asks Claude a question, the server writes the full payload to disk and prints a single-line pointer to stdout. Claude Code delivers that line as a notification; Claude reads the file and — for prompts — answers via the respond_to_prompt MCP tool.
• Ships one slash command, /openconclave:open, that opens the editor in your default browser.

Plugin storage

Data lives at ~/.claude/plugins/data/openconclave-openconclave/, per Anthropic's plugin data guidance. This path:

• Survives plugin updates — your conclaves, runs, KBs, and session history don't move when the plugin version bumps.
• Is auto-deleted on full uninstall unless you pass --keep-data to claude plugin uninstall.

First-run migration: if you previously installed the standalone oc CLI and have data at ~/.openconclave/, the plugin's SessionStart hook copies openconclave.db, sessions/, outputs/, and instructions/ into the plugin data dir the first time. The legacy directory is left in place — downgrading back to the standalone CLI still works.

Uninstall

claude plugin uninstall openconclave@openconclave

By default this also removes the plugin data dir (conclaves, runs, KBs). Pass --keep-data to preserve it if you plan to reinstall.

claude plugin uninstall openconclave@openconclave --keep-data

Troubleshooting

• Port 4000 in use — the monitor tries to detect an existing OC server and attaches instead of binding a second one. If you still see EADDRINUSE, another process owns :4000; stop it and /reload-plugins.
• UI is 404 — the client bundle didn't build. The SessionStart hook runs bun install + bun run --filter client build on the first session; if it fails (no Bun, offline), the server runs but has no UI. Check stderr.
• /doctor shows CLAUDE_PLUGIN_ROOT missing — you have a stale root .mcp.json from an older plugin version. Remove it; MCP config lives inline in .claude-plugin/plugin.json now.
• UI / MCP died after closing Claude Code — on Windows, Claude Code's process tree teardown can leave an orphan bun process holding :4000. The plugin now writes a pidfile at ${CLAUDE_PLUGIN_DATA}/oc.pid and captures stderr at ${CLAUDE_PLUGIN_DATA}/server.log so restarts detect and reap stale servers. If something still misbehaves, cat that log.

Security & trust model

Installing the OpenConclave plugin is a real trust decision. Here's what it actually does so you can decide whether you're OK with it.

What the plugin's own code does

• Starts a local HTTP + MCP server on localhost:4000. Never listens on a public interface, never binds 0.0.0.0 without you asking.
• Serves the editor UI and the API from the plugin install directory.
• Writes its database, per-run session artifacts, and logs under ~/.claude/plugins/data/openconclave-openconclave/.
• On first install, copies an existing ~/.openconclave/ directory into the plugin data dir (migration). The source is not modified or deleted.
• Does not phone home. No telemetry, no outbound connection from the plugin's own code. Landing-page manifest polls from your machine for update notifications are opt-in under Settings → Advanced.

What conclave agents can do

A conclave is a graph you (or someone whose graph you installed from the marketplace) build in the editor. Agents in a conclave can use whichever tools their node is configured with — bash, read/write/edit, web fetch, knowledge search, user-registered MCP tools, etc. These tools run locally with your shell's privileges.

• Tool access is per-node and opt-in. A new agent node starts with no tools; you add them.
• Conclaves imported from the marketplace ship their own tool wiring. Review the graph before running anything you didn't build yourself — same posture as running a shell script you pulled off the internet.
• Conclave runs happen in a worktree when you enable Git isolation on the trigger, so mutating pipelines can't touch your main branch without your opt-in.

What providers the plugin talks to

Only the ones you configure. OC has no default provider credentials. You supply:

• An ANTHROPIC_API_KEY or OPENAI_API_KEY or any OpenAI-compatible provider URL + token, stored in OC's settings DB (not in environment, not shipped to Anthropic/OpenAI beyond the model API calls you make).
• Optionally, a local Ollama endpoint for offline-only runs.

Keys live in ~/.claude/plugins/data/openconclave-openconclave/openconclave.db. They're never sent to Anthropic, Anthropic's plugin registry, openconclave.com, or anywhere else by the plugin itself.

What happens on uninstall

claude plugin uninstall openconclave@openconclave

Removes the plugin code AND the plugin data dir by default. That deletes your conclaves, runs, KBs, and API keys stored via OC.

claude plugin uninstall openconclave@openconclave --keep-data

Removes the plugin code but keeps ~/.claude/plugins/data/openconclave-openconclave/ intact. Reinstall picks up where you left off.

Trust boundary, in one line

The plugin inherits your shell's permissions; it does not escalate them. Everything it can do, you could do yourself by running oc standalone. The shape of the trust decision is "do I trust the OC server to run the graphs I build with the tools I wire up" — the same decision you make when writing a Makefile.

Manual install

git clone https://github.com/openconclave/oc.git
cd oc && bun install && bun start

Docs

Architecture · Conclave composition · Security · Lab journal

License

MIT