by Hoylon
Local-first, auditable multi-agent control room for coding, review, evidence, and private remote work.
# Add to your Claude Code skills
git clone https://github.com/Hoylon/peerbridge-mcpGuides for using ai agents skills like peerbridge-mcp.
Last scanned: 8/26/2026
{
"issues": [],
"status": "PASSED",
"scannedAt": "2026-08-26T04:38:41.513Z",
"npmAuditRan": true,
"pipAuditRan": true,
"promptInjectionRan": true
}peerbridge-mcp is an open-source ai agents skill for AI coding assistants such as Claude Code, Codex CLI, and ChatGPT, built by Hoylon. Local-first, auditable multi-agent control room for coding, review, evidence, and private remote work. It has 238 GitHub stars.
Yes. peerbridge-mcp passed SkillsLLM's automated security scan — a dependency vulnerability audit plus prompt-injection heuristics — with no high-severity issues. You can read the full report in the Security Report section on this page.
Clone the repository with "git clone https://github.com/Hoylon/peerbridge-mcp" and add it to your Claude Code skills directory (see the Installation section above).
peerbridge-mcp is primarily written in Python. It is open-source under Hoylon on GitHub, so you can review or fork the full source.
Yes. SkillsLLM lists many other AI Agents skills you can browse and compare side by side. Open the AI Agents category from the badge at the top of this page, or use the Related Skills and comparison links further down to weigh peerbridge-mcp against similar tools.
No comments yet. Be the first to share your thoughts!
⚠️ Third-Party Software Notice
This skill is third-party open-source software developed and hosted independently on GitHub. SkillsLLM is an informational directory and does not control or maintain the underlying repository.
Any security checks, ratings, or warnings displayed by SkillsLLM are automated and limited in scope. They do not constitute a security certification or guarantee that the software is safe, error-free, or free from malicious code, vulnerabilities, compromised dependencies, or prompt-injection risks.
Review the source code, permissions, dependencies, and configuration before installing or running any third-party skill. Use is at your own risk. To the maximum extent permitted by applicable law, SkillsLLM is not liable for losses arising from third-party software.
PeerBridgeControlRoom-0.1.0a6-windows-x64-portable.zip from the
Alpha 6 release.SHA256SUMS.txt on the release page.Launch PeerBridge.cmd.The portable build does not install a service, modify an existing Python environment, or
include provider credentials. Local workspace data stays under
%LOCALAPPDATA%\PeerBridge\workspace. Prefer source installation? Continue to
Install from source.
PeerBridge is more than a collection of chat tabs. Its provider-native adapter layer keeps Codex JSON-RPC, Claude Code stream-json, Grok/Kimi ACP, provider APIs, and local runtimes as separate capability contracts. Every Agent is bound to a stable identity, selected model, permission policy, room context, task, and evidence trail.
Use it for parallel implementation and review, bounded technical debate, task ownership, approved shared memory, mutual scoring, cross-agent audits, and human-controlled release work. PeerBridge runs locally on SQLite, keeps provider credentials out of chat and project history, and preserves a SHA-linked record of messages, decisions, evidence, scores, permissions, and handoffs.
Human or Agent messages can wake the room, collaboration stops on consensus, blockers, stagnation, or explicit limits, and the operator can intervene at any time.
| What maintainers get | PeerBridge behavior |
|---|---|
| Provider-native adapter layer | Codex, Claude Code, Grok, Kimi, provider APIs, compatible endpoints, and local runtimes keep separate audited identities and capability boundaries. |
| Work instead of reply cascades | Parallel implementation, review, discussion, and release workflows stop on consensus, blockers, stagnation, or explicit limits. |
| Governed write access | Permission cards, approved isolated Git worktrees, source-state checks, diffs, and proof records stay human-controlled. |
| Durable project context | Selected history imports, room memory, task briefings, and handoffs are source-bound instead of silently merged. |
| Reviewable decisions | Agent activity, answers, evidence, mutual scores, audits, token usage, and stale-proof warnings remain visible. |
| Local ownership | SQLite data and credentials stay on the operator's machine; browser remote control remains private, while the allow-listed Tailcat companion is visible and enabled by default. |
flowchart LR
A["Provider-native adapters"] --> C["Per-Agent capability union"]
C --> P["Capability-aware approval broker"]
P --> W["Governed worktree and writer lease"]
W --> R["Parallel review and bounded collaboration"]
R --> E["SHA-linked evidence and Proof Bundle"]
H["Human operator"] --> P
H --> R
Run the provider-free maintainer showcase with one command:
python examples\demo_workflow.py --workspace demo-workspace --scope demo
Its public receipt proves that an overlapping second writer was rejected, two independent reviewers satisfied quorum, completion rehashed the synthetic artifact, and the audit chain verified with zero writes. It contains no provider credential or lease capability. See the claim-to-test evidence map.
Status: alpha. The coordination and audit core is tested, but the public API and database schema may change before 1.0.
Operational continuity and commercialization are explicit design boundaries. See memory and long-running operations and the open-core boundary. Tailscale Serve remains the private browser transport; Tailcat CLI connectivity is visible and enabled by default but remains independently allow-listed. Roadmap entries are not capability claims.
Running two coding agents against one repository creates predictable failure modes:
PeerBridge addresses those coordination failures without making either agent the boss. One task has one writer lease, peers can review each other, and a human can intervene.
flowchart LR
H["Human operator"] --> M["Local control room"]
C["Codex / Claude Code"] --> S1["PeerBridge stdio process"]
A["Grok / Kimi / DeepSeek / Gemini / local Agents"] --> S2["PeerBridge stdio process"]
M --> S3["PeerBridge stdio process"]
R["Private mobile UI"] -->|"Tailnet HTTPS"| T["Tailscale Serve"]
T -->|"loopback only"| S4["Human MCP message gateway"]
S1 --> DB[("Shared SQLite store")]
S2 --> DB
S3 --> DB
S4 --> DB
DB --> L["SHA-linked audit events"]
DB --> W["Mailbox, leases, reviews, proofs"]
Each MCP client launches its own stdio server process with a distinct --agent-id.
Those processes coordinate through one project-local .peerbridge/peerbridge.sqlite3
database. SQLite WAL mode and BEGIN IMMEDIATE transactions serialize state changes.
PeerBridge is one local desktop product, not a hosted web service. Both desktop surfaces use the same local Python, SQLite, and MCP core:
A fresh workspace shows a local two-preview chooser before the Control Room opens. The
selection is stored in .peerbridge/ui-preferences.json; later launches open that surface
directly until the operator changes Appearance and restarts.
grok sessions list, and Claude/Kimi use their documented local session records. The UI
lists bounded metadata first, selects nothing by default, and reads only conversations the
operator checks; JSON/JSONL file import remains available for every provider.