MultiContext Chat

Parallel isolated LLM chats with per-chat prompts/tools, queued cross-chat messaging, and optional response compression. LibreChat supplies the Agent runtime, model/provider integrations, Web Search, MCP, code execution, knowledge, and other existing tools.

Release status: macOS internal verification build. Core workspace, queue, Agent selection, retry/stop, Compile, and external MCP flows are usable, but one locked-screen UI pass remains in the runtime record. Do not distribute the unsigned local DMG; Developer ID signing and notarization are required.

Core behavior

• One human prompt is broadcast to every active chat.
• Each chat keeps an independent context and an independent FIFO input queue.
• Different chats run in parallel; one chat processes only one queued prompt at a time.
• Workspace system prompt is shared and editable.
• Each chat has an editable developer prompt.
• Human broadcasts and cross-chat queued prompts are sent as normal user input.
• Chats do not automatically receive another chat's history or tool results.
• Optional tools let a chat list peers, inspect selected peer messages, or queue a prompt to one or two peers.
• The application does not decide what an input means or when a model should use those tools. Prompt/model behavior owns that decision.
• SETTLED is a runtime state: no active generation, queued work, or blocked failed turn remains.
• Compile is manual. Its result is user-facing only and is not injected back into member contexts.

LibreChat modes

native (default)

Designed for gpt-oss when exact role separation and LibreChat-owned conversation context are required. Apply the small, idempotent host patch:

node scripts/patch-librechat.mjs /path/to/LibreChat

Rebuild LibreChat, then set:

MULTICONTEXT_LIBRECHAT_MODE=native

Native mode stores one stable LibreChat conversation id per MultiContext member and continues it with previous_response_id. Request-level cross-chat tools are exposed to the model but executed by MultiContext, not LibreChat. After a model function_call, the native continuation re-sends the answered function_call item followed by its function_call_output, together with previous_response_id and the same tool definitions. It does not replay system, developer, user, or local history. This explicit tool round trip is required because LibreChat persistence alone does not retain enough structured tool-call state for gpt-oss to ground a dangling function_call_output. Native mode fails fast if the patched conversation-id header is absent. See docs/GPT_OSS.md.

compat

Works with stock LibreChat. MultiContext owns each member's bounded history, sends store:false, and replays that member history on every turn. The request boundary still contains separate system and developer items, but stock LibreChat may normalize them internally.

Quick start

cp .env.example .env
# configure LIBRECHAT_BASE_URL and LIBRECHAT_API_KEY
npm run check
npm start

Open http://127.0.0.1:4317.

Docker:

docker compose up --build

If LibreChat must call the generated cross-chat Actions, configure MULTICONTEXT_PUBLIC_URL to a URL reachable from the LibreChat process/container.

Desktop (macOS)

Native launcher you can double-click from Finder/Dock. It bundles the MultiContext Node server (no dependency on the Git checkout) and can start LibreChat and your model backend for you.

npm run desktop:dev   # Tauri dev
npm run desktop:build # production MultiContext.app
open src-tauri/target/release/bundle/macos/MultiContext.app

Product usage and recovery

1. Launch MultiContext.app; it reuses healthy external services or starts the configured managed LibreChat/model services.
2. On first launch, save the LibreChat Remote Agents connection key in Settings. It is validated against the Remote Agents API and stored only in macOS Keychain, never in config.json or logs.
3. Choose an Agent explicitly in safe mode, or use auto_first when first-Agent selection is acceptable. Safe mode rejects an unconfigured broadcast with AGENT_SELECTION_REQUIRED.
4. When creating a workspace, choose 1–4 initial chats (2 is recommended for parallel comparison). Assign an Agent to each chat before broadcasting.
5. On service/workspace refresh failure, use Retry / 再確認. The last saved view remains visible, retry progress is shown, and privacy-safe diagnostics can be copied from the AI Stack status dialog.
6. Archive/delete only inactive workspaces. Deletion requires typing the exact workspace name; state saves retain a .bak recovery copy.

The product provides isolated chat contexts with per-chat FIFO queues and parallel execution. Compile is allowed only at SETTLED and does not write back into member histories. REST and external MCP share the same application operation layer, while LibreChat Agent tool use remains a separate path.

External Control MCP

MultiContext exposes itself as an MCP server so external clients (OpenCode, Cursor, Claude Code) can operate workspaces programmatically. This is not LibreChat Agent MCP.

• Endpoint: http://127.0.0.1:4317/mcp (Streamable HTTP, loopback by default)
• Auth: MULTICONTEXT_MCP_TOKEN (distinct from LibreChat key, stored in macOS Keychain when using Desktop)
• OpenCode config: Desktop Settings → 外部連携 — MCP Server → OpenCode設定をコピー generates a ready opencode.json remote entry (see docs/MCP.md).

# enable and copy config via Desktop UI, or manually:
MULTICONTEXT_MCP_TOKEN=$(openssl rand -hex 32) npm start

See docs/MCP.md for tool list, agent selection, SETTLED/Compile semantics, and security.

After the first launch, just open MultiContext.app — it detects healthy external GPT-OSS/LibreChat or starts managed ones (new installs default to managed) and opens the existing UI automatically once everything is ready. On first run, open Settings and set your LibreChat directory, llama-server, GPT-OSS model, chat template, and the LibreChat Remote Agents connection key (validated against the Remote Agents API, not normal user auth — saved once in macOS Keychain via the LibreChat 接続 section — no Terminal, no config.json, empty field preserves the stored key, deletion requires explicit 保存済みキーを削除, and the key is never shown back). 保存して開始 saves a newly entered key together with config in one step; Retry always performs a fresh readiness attempt. See docs/DESKTOP.md for first-run setup (external vs managed services, Keychain credential storage, logs at ~/Library/Logs/com.unjuno.multicontext/, and Gatekeeper notes for unsigned builds).

Runtime states

• RUNNING — at least one member is generating.
• PENDING — queued work is waiting and can run.
• BLOCKED — a model/tool request failed; the prompt was returned to the front of that member FIFO and requires explicit Retry.
• SETTLED — all active member queues are empty and no member is running or blocked.

Stop aborts active requests and clears pending work. An interrupted process does not silently lose an in-flight prompt: startup recovery requeues it.

Completed or blocked workspaces can be archived without deleting their history. Archived workspaces are hidden from the default list and can be found through the アーカイブ済み filter and restored. Running or queued workspaces cannot be archived, so active work is never hidden accidentally.

Cross-chat tools

Each member exposes an OpenAPI Action URL. The Action contains:

• list_chats()
• inspect_chat(target, query, limit)
• send_to_chat(targets, prompt) where targets contains one or two chat ids/exact names

Tool availability is controlled in MultiContext and in LibreChat Agent configuration. Message interpretation and tool-use policy remain prompt/model responsibilities.

Compile

Compile is available only when the workspace is SETTLED and only runs when the user presses Compile. compilePrompt is editable. The compiler receives bounded recent visible records from active chats and returns one compressed response. No compile output is written into member histories.

Validation

npm run check runs syntax checks and Node tests covering context/queue isolation, FIFO + cross-member concurrency, failure requeue and explicit Retry, persisted in-flight recovery, Stop stale-result suppression, native/compat request shaping, LibreChat health/agent discovery, cross-chat Action delivery, HTTP runtime state/Compile gating, and patch idempotency.

Run npm run smoke against a real LibreChat deployment. Set MULTICONTEXT_SMOKE_AGENT_ID to include real generation and native thread-continuation checks.

License

MIT. LibreChat remains a separate MIT-licensed upstream dependency.