> **Machine-Readable Markdown Mirror** · Canonical HTML: https://senrelay.com/
> **Product**: Senrelay (续境 · `qingsongcui/ContextCore`) · **Version**: `v1.0.0-beta.4` · **Updated**: `2026-11-20`
> **Pricing & License**: **$0 (100% Free & Open Source under the MIT License, no signup)** · [Pricing & Limits](https://senrelay.com/pricing.md) · [Extended Spec](https://senrelay.com/llms-full.txt)

CONTEXT IS CONTINUITY

# Switch agents, keep the thread .

Switch tools. Keep your thinking.

**Senrelay** is a local multi-agent context workbench for macOS 14.0+ (Apple Silicon), free and open source under the MIT License ($0, no account required). Type `/contextcore` or say "Relay" across Claude Code, Codex, Cursor, Antigravity, OMP, and OpenCode to continue recent conversations, capture source-linked memories in chat, and manage providers, MCP, and Skills in one place.

[Download macOS Beta.4](https://github.com/qingsongcui/ContextCore/releases/download/v1.0.0-beta.4/Senrelay-macOS.zip)[See how relay works](https://senrelay.com/#handoff)[Watch 42s film](https://senrelay.com/#launch-film)

Multi-host configuration management Implicit auto-memory Cross-agent relay MIT open source

Built for multi-agent work LOCAL FIRST. CONTEXT ALWAYS. [↓](https://senrelay.com/#why)

### Watch the agent handoff 0:42

Your browser cannot play inline MP4 video. [Download the 1080p launch film](https://github.com/qingsongcui/ContextCore/releases/download/v1.0.0-beta.2/senrelay-keynote-en.mp4).

42s · 1080p · English narration
[Download EN MP4](https://github.com/qingsongcui/ContextCore/releases/download/v1.0.0-beta.2/senrelay-keynote-en.mp4)[EN Captions (.srt)](https://github.com/qingsongcui/ContextCore/releases/download/v1.0.0-beta.2/senrelay-keynote-en.srt)[ZH Cut (.mp4)](https://github.com/qingsongcui/ContextCore/releases/download/v1.0.0-beta.2/senrelay-keynote-zh.mp4)

01 / THE REASON Keep context with the project

## Spend time moving forward, not re-explaining.

Architecture settled, code halfway done. New agent, another briefing from scratch?

Senrelay brings **full-transcript cross-agent relay , source-linked companion auto-memory , and multi-host MCP/Skills management** into one local Mac workbench. Companion tools track working task state and save source-linked memories as you work; in a new assistant, type `/contextcore` or say "Relay" to pick from a short brief list, read the full transcript, and compact in context to continue (see the [workflows guide](https://senrelay.com/guide.html#everyday-use)).

[See how it works](https://senrelay.com/#handoff)

[ 01 ]

### Provider and model switching

Switch hosts from the top bar, fill provider presets in one click, test latency, switch or sync MCP and Skills across hosts, and import existing local configurations on demand.

[ 02 ]

### Companion Implicit Auto-Memory

Once connected, companion MCP tools and Skills can save structured memories with links to their source. You can also edit memories or import documents.

[ 03 ]

### One-phrase cross-agent relay

Type /contextcore or say "Relay" in a connected core host to pick from a short-brief list without calling a model; once selected, it pages through the full transcript and compacts everything in context—brief first on number-only selection, or continue immediately with a task.

Built for everyday multi-agent workflows

- **Cross-Host Coding Relay** Carry change boundaries and test status between connected core hosts.
- **Research & Architecture** Carry multi-turn sources, ruled-out options, and constraints into another model to keep reasoning.
- **Implicit Auto-Memory** Save coding rules, preferences, and lessons so a later session can read them.
- **Unified Configuration Control** Switch and sync providers, MCP, Skills, and system prompts in one click with automatic rollback snapshots and external-edit conflict diff prompts.

02 / AGENT MANAGEMENT Models, tools, and preferences in their own place

## Your agents. Unified here.

Six Agent workspaces manage providers, models, MCP, Skills, and rules. CLI and desktop entries group by the configuration they share. Preview the files before applying changes and use the change history to restore a previous configuration.

[AI coding agent management](https://senrelay.com/agent-management.html) · [Claude Code, Cursor, and Codex MCP config paths](https://senrelay.com/hosts.html) · [Claude Code and Codex usage and quotas](https://senrelay.com/ai-coding-usage-tracker.html)

- 01 / SWITCH HOST Select target host in top bar
- 02 / PRESET & SPEEDTEST Apply presets & test latency
- 03 / SYNC RESOURCES Sync MCP, Skills & instructions
- 04 / GUARD & APPLY Conflict check & safe rollback

### Provider Presets & Local Config Sync

Start from a provider preset, manual definition, or native import. Review the model, protocol, and account before applying to a supported assistant.

### Multi-Mode Routing & Keychain

API keys stay isolated in macOS Keychain. Choose direct or local-proxy mode where the assistant supports it. Routing supports model aliases and ordered fallback providers on supported hosts.

### MCP Tools & Companion Skills

Configure stdio or HTTP tool servers and import Skills from local folders, ZIPs, or GitHub. Connects hosts to local `http://127.0.0.1:8420/mcp/` under the server key `contextcore` alongside the companion `/contextcore` relay skill (never run unrelated PyPI `contextcore`).

### System Prompts & Project Bundles

Maintain coding conventions and global rules in one place, or bundle project-specific providers, models, MCP tools, Skills, and prompts to preview diffs and switch in one click.

### Request Traces, Usage & Latency

Review observed token counts, retries, and estimated costs. Check supported account quotas separately.

### External Edit Guard & Safe Snapshots

Clearly distinguishes applied state, pending updates, and external manual edits. Every write saves a recovery snapshot with diff inspection, conflict alerts, and instant rollback—and startup or config import never silently overwrites installed host Skills.

Each host uses its own native configuration format and capability scope; the workbench provides live write verification, latency checks, and environment conflict alerts to keep every change predictable.

[See assistant support](https://senrelay.com/#connections)

03 / CROSS-AGENT PROJECT RELAY Context in place, thinking in place

## Switch agents. Pick up the entire thread.

Type /contextcore or say "Relay" in a connected core host to list recent cross-host conversations with short recognizable briefs without calling a model; after you pick one, instead of truncating into an external handoff card, the receiving agent pages through the full transcript and compacts everything in context—reply with just a number to report and wait, or include a task to continue right away.

[See which relay paths were verified](https://senrelay.com/relay-workflow.html#relay-evidence)

SHORT BRIEF LIST · FULL TRANSCRIPT COMPACTION

### What full-transcript relay preserves.

Identify the session from short briefs → read every message from offset 0 within safe page budgets → compact the entire conversation directly in the receiving agent with these six guarantees:

STEP 01 · TRIGGER RELAY **Type /contextcore or "Relay" in the connected host** Reads local native session indexes to list candidates without extra model calls

STEP 02 · PICK BY SHORT BRIEF **View index, source agent, timestamp, title & short brief** Preserves native session IDs and worktree/directory identity so same-title or unassigned sessions stay distinct

STEP 03 · PAGE FULL TRANSCRIPT & COMPACT **Read full transcript pages sequentially and compact in context** 48KB page cap guards tool budgets; number-only selection reports in read-only stand-by, while explicit tasks execute

- #### Short brief first, zero pre-summarization Lists show only recognizable titles and short briefs without running a background summarizer.

- #### Native session & worktree identity Retains raw session UUIDs and worktree paths so same-title or unassigned chats never collide.

- #### Paged full-transcript reading Pages from offset 0 through the final turn with strict byte bounds to avoid MCP response overflows.

- #### In-context compaction, no handoff cards The receiving agent compacts the full read transcript in its own context instead of relying on lossy handoff cards.

- #### Working tasks & fail-closed constraints Carries active working tasks and locked rules; protected constraints fail closed when over budget instead of silent clipping.

- #### Read-only stand-by & on-demand refresh Number- or ID-only selection locks must_not_execute=true to report and wait; refreshes only the chosen source.

**Pick a number to read & report first, or add a task to act.**

Replying with only a number or session ID reads the full transcript, compacts it into a clean briefing, and waits without editing files; adding an explicit task continues right away with full context.

04 / AUTO-MEMORY & WORKSPACE Belongs to the project, not one chat

## The story of the project, always within reach.

01 — PROJECT

### Scattered chats, back in one project.

Stop digging through scattered chat windows. Create directory-free projects up front, and let chats across agents and worktrees, working task state, decisions, and references converge in one project with direct source links.

Organized by project · directory-free creation · cross-source search

02 — AUTO-MEMORY

### Let experience accumulate, not chats pile up.

No manual note-taking required—companion MCP tools and Skills maintain working task progress alongside six categories of version-governed long-term memories with provenance-checked lesson promotions, supersession trails, pinned rules, and conflict review.

Working tasks + versioned memory · Source-traceable

03 — KNOWLEDGE

### Keep your documents working for you.

Import requirements, guidelines, and references for authorized assistants to search. Review the source and extracted text in the workspace.

Markdown · TXT · JSON · PDF · DOCX · DOC

04 — FIND WITH SOURCES

### Ask in your own words. Check the original.

Use a question or keyword to search chats, memories, and documents in the current project or across all projects. Open any match to inspect its exact source and original transcript before handing it to an assistant.

### Optional enhancement: local semantic search & knowledge graph

Local vector search finds conceptually related discussions, while source-backed entity graphs connect clues across sessions. Both run locally on your Mac and can be enabled on demand while keeping keyword search instant.

05 / START WITH ONE TASK One assistant, one task to start

## Ready in four steps. Start relaying.

Launch the local service and workbench to unify multi-host configurations and companion components; capture source-linked project memories during daily chats and relay across agents with a single command.

- 01 ### Start Local Workbench Open the desktop app or follow the [local setup guide](https://senrelay.com/guide.html#download-requirements) to start the local service, and optionally preview and sync existing local configurations.
- 02 ### One-Click Connect & Sync Switch hosts in Agent Management, click One-Click Connect to wire history paths, MCP, and /contextcore Skills, and enable providers with live latency checks.
- 03 ### Capture Auto-Memories Work normally in your favorite agent while companion MCP tools and Skills save source-linked structured memories with a concise receipt.
- 04 ### Type "Relay" in New Agent In a new agent, type /contextcore or say "Relay", pick a session from the short-brief list, then page through and compact the full transcript in context to stand by or continue (see [cross-agent relay workflow](https://senrelay.com/relay-workflow.html)).

### **Local first, clear permissions.** Understand data and permission boundaries

All session indexes, working tasks, project memories, and config snapshots stay on your Mac, with API keys isolated in Keychain. Reading chat history, editing host configs, and downloading optional semantic models are confirmed separately; protected rules fail closed instead of silent clipping when over budget, and search never connects to the network.

06 / IN YOUR OWN TOOLS Keep your way of working

## Your familiar agents. Connected to the same project.

Keep using familiar tools. Senrelay stores and provides shared project context.

- **Codex** Desktop / CLI Carry progress into the next stretch of work
- **Cursor** IDE Pick up existing thinking in the editor
- **OMP** Oh My Pi Let context follow your workflow
- **Antigravity** Desktop Change environments, keep the project background
- **Claude Code** CLI Continue the previous stretch in the terminal
- **OpenCode** CLI Bring project context into terminal workflows

The six core agents above support full session capture, implicit auto-memory, and cross-agent relay; additional extended hosts support configuration file management only.

**Host Support & Local Runtime Boundaries**

| Tier | Covered Hosts / Runtime | Actual Capabilities |
| --- | --- | --- |
| 6 Core Context Hosts | Claude Code, Codex, Cursor, Antigravity, OMP, OpenCode | Native session capture with worktree disambiguation, working tasks and versioned auto-memory, `/contextcore` short-brief selection and paged full-transcript compaction (number-only sets `must_not_execute=true`), plus `AGENTS.md` preview and config sync (see [all 15 host config paths](https://senrelay.com/hosts.html)). |
| 9 Extended Config Hosts | Claude Desktop, Gemini CLI, Grok Build, Kimi Code, MiniMax Code, Pi, Hermes, OpenClaw, agy | Configuration diff preview, writes, and rollback snapshots only (JSON / JSON5 / TOML / YAML); no session capture or conversation relay. |
| Local Architecture | `macOS 14.0+` · `Apple Silicon` (`arm64`) | Loopback only on `127.0.0.1:8420` (dev `18420`) with MCP server key `contextcore` (`/mcp/`); local `SQLite` + `FTS5` (`0600`) with `contextcore backup` online backups and API keys in macOS Keychain. |

**Release, Pricing & Known Limitations (Verified 2026-11-15)**

| Item | Verified Fact | Notes & Constraints |
| --- | --- | --- |
| Pricing & License | **$0 · MIT Free & Open Source** · No Signup | Permanently free with no paid tiers or cloud paywalls; model API keys stay with your chosen providers and macOS Keychain. |
| Release & Install | `v1.0.0-beta.4` · Public Beta | [Download Senrelay-macOS.zip](https://github.com/qingsongcui/ContextCore/releases/download/v1.0.0-beta.4/Senrelay-macOS.zip) or [build from source](https://senrelay.com/guide.html#download-requirements). Do not run unrelated `pip install contextcore`. |
| Where Not to Use | macOS 14.0+ arm64 only · Single-Mac · First-launch clear | Not supported on Windows, Linux, or Intel (`x86_64`) Macs, and has no multi-device cloud sync; Beta builds are not Apple-notarized and require a one-time `xattr -cr /Applications/Senrelay.app` ([5-option comparison](https://senrelay.com/compare.html)). |

### Multi-host integration & capability coordination

The six core hosts—Codex, Claude Code, Cursor, Antigravity, OMP, and OpenCode—support full local session capture, companion implicit auto-memory, and cross-agent conversation relay. Additional extended hosts support previewed configuration file management only.

Configuration file writes, environment variable conflict checks, endpoint latency tests, and live process reload verification are tracked independently so status stays transparent.

07 / A FEW GOOD QUESTIONS

## A few more questions.

About handoffs, data, and what you can do today.

### What is Senrelay (ContextCore), and how does it differ from CLAUDE.md or cloud memory MCPs?

Senrelay (open-source repository `ContextCore`, registered in host MCP configs as `contextcore`) is a free ($0 MIT, no signup) macOS 14.0+ Apple Silicon local workbench. Unlike static `CLAUDE.md` or `AGENTS.md` repo instructions or hosted cloud memory MCPs, it keeps everything on your Mac while combining `/contextcore` full-transcript cross-agent relay, versioned shared memory, and multi-host MCP/Skills sync.

### How does Cross-Agent Conversation Relay work in practice?

Type `/contextcore` or say "Relay" in the connected host to list numbered recent cross-host sessions with short recognizable briefs—zero model calls during discovery. Once you select a conversation, the receiving agent pages through the full transcript via `read_conversation_messages` (capped per MCP page to protect context budgets) and compacts the entire thread directly in context: replying with just a number locks `must_not_execute=true` to report first and wait; attaching a task continues immediately with full context.

### When does companion memory get saved?

Once connected, an assistant can call the companion memory tool to save source-linked content. Check the workbench for the saved record and its source. Senrelay maintains dual-tier working memory combining working tasks (`working_tasks`) and version-governed long-term memories: superseding a decision marks the prior record outdated instead of overwriting history, ambiguous changes route to Pending Conflicts, and lesson promotions require source-dialogue provenance alongside rule pinning and document imports.

### Which local configurations can I sync or import?

The six Agent workspaces can import and manage supported native configurations. Preview changes to providers, models, MCP, Skills, or rules, and export Senrelay's own backup. Check each assistant's supported operations and file paths.

### Will changing one assistant overwrite my other assistants?

Writes follow the selected native files. Entries that share a file, such as Codex CLI and Desktop on the default profile, share the change. Preview the affected files before applying.

### Do my transcripts and API keys stay safe on my Mac?

Yes—100% of your transcripts, memories, and config snapshots remain on your local Mac and are never uploaded to the cloud. All data lives in your local SQLite database (`0600` file permissions), and API keys are isolated in macOS Keychain and excluded from config exports by default.

### How much does Senrelay cost? Do I need an account or subscription?

Senrelay is 100% free ($0) under the open-source MIT License, with no account registration, no subscription tiers, and no cloud paywall. Every cross-agent relay, auto-memory, and multi-host configuration feature runs locally; you only provide your own model provider API keys in your AI coding assistants.

### What are Senrelay's real limitations, and when should I NOT use it?

Do not use Senrelay if you work on Windows, Linux, or Intel Macs, or if you require multi-device cloud sync and hosted team collaboration. Current release `v1.0.0-beta.4` runs exclusively on single-machine macOS 14.0+ (Apple Silicon `arm64`); the 9 extended hosts support configuration file management only; and Public Beta builds require a one-time `xattr -cr /Applications/Senrelay.app` command before Apple notarization is added. Compare tradeoffs in our [5-option comparison & stack picker](https://senrelay.com/compare.html#picker).

### How do I install and launch the workbench right now?

[Download macOS Beta.4](https://github.com/qingsongcui/ContextCore/releases/download/v1.0.0-beta.4/Senrelay-macOS.zip) and follow the [installation guide](https://senrelay.com/guide.html#download-requirements) to verify the checksum and open the app. You can also follow the [local setup guide](https://senrelay.com/guide.html#launch-workbench) to run from source on loopback port `127.0.0.1:8420` / `18420`. Note: never run `pip install contextcore`—the PyPI package of that name is an unrelated third-party project.

[See the start guide](https://senrelay.com/#start)

KEEP YOUR CONTEXT. KEEP CREATING.

## Next time, start from here .

Launch your local multi-agent workbench in minutes. Manage multi-host configurations, capture source-linked memories as you chat, and relay conversations across agents anytime.

[Download macOS Beta.4](https://github.com/qingsongcui/ContextCore/releases/download/v1.0.0-beta.4/Senrelay-macOS.zip)[Read the local setup guide](https://senrelay.com/guide.html)[15-host config paths](https://senrelay.com/hosts.html)[Cross-agent relay workflow](https://senrelay.com/relay-workflow.html)[Compare 5 approaches](https://senrelay.com/compare.html)[Visit source repository](https://github.com/qingsongcui/ContextCore)

**Author & Open-Source Repository**

Senrelay (open-source repository [qingsongcui/ContextCore](https://github.com/qingsongcui/ContextCore)) is created and maintained by independent developer **Qingsong Cui (George O'Nair)** ([GitHub @qingsongcui](https://github.com/qingsongcui) · [X @George_ONair](https://x.com/George_ONair))—unrelated to third-party `context-core.dev` or PyPI `contextcore`.

**Local Privacy & License**

Strictly binds to local loopback `127.0.0.1:8420` with zero telemetry, zero uploads, and no signup; SQLite databases (`0600`) and Keychain credentials stay on your Mac. Provided "AS IS" under the [MIT License](https://github.com/qingsongcui/ContextCore/blob/main/LICENSE).

macOS 14.0+ (Apple Silicon) · `v1.0.0-beta.4` · **$0 MIT open source** · Released 2026-11-19
