Companion Connector MCP

Local Companion MCP app for ChatGPT New App custom server connection.

V5 capabilities

This version is built for large local work, Fable handoff, screenshots/images, URL reading, media metadata, and future 21-service routing.

Core MCP tools

• search

• fetch

• register_file_pointer

• read_file_slice

• read_result_slice

• create_summary_job

• create_file_digest_job

• create_directory_inventory_job

• register_image_pointer

• ingest_image_base64

• get_image_data

• create_image_inspection_job

• ingest_chat_transcript

• ingest_attachment_base64

• create_fable_bundle

• run_fable_bundle

Fable and batch tools

• create_question_batch

• run_question_batch

• ask_fable_big

• create_fable_improvement_survey

• connector_health_report

• list_jobs

• list_fable_runs

• create_fable_prompt_file

• get_job_status

• list_registered_resources

URL/media/handoff tools

• fetch_url_text

• extract_links_from_url

• create_url_snapshot_job

• register_video_pointer

• create_media_metadata_job

• create_handoff_queue_item

• list_handoff_queue

• list_mcp_services

• describe_mcp_service

Related MCP server: ChatGPT2LocalBridge

What it solves

• Big files are referenced by path instead of pasted into chat.

• Large questions can be written into local prompt bundles for Fable.

• Multiple questions can be batched into one file-backed request.

• URLs can be fetched into bounded snapshots.

• Links can be extracted from a page.

• Attachments can be stored as connector resources.

• Screenshots/images can be registered by path or small base64 payload.

• Media/video files can be registered and inspected with ffprobe when available.

• Fable can be given a bundle prompt file with transcript, attachment pointers, full text sections, image metadata, media metadata, and hashes.

• 21 planned MCP service folders are exposed as a service catalog.

• Outputs are written to results/ and read by slice.

Start locally

npm start

Default endpoint:

http://127.0.0.1:8788/mcp

Tests

npm test
npm run test:fable
npm run test:v4
npm run test:v5

Hidden PowerShell 7 startup

.\scripts\start-hidden.ps1

Stop by port:

.\scripts\stop-companion.ps1

Tunnel setup

See TUNNEL_SETUP.md.

Safety boundaries

• Reads are restricted to configured allowed roots.

• Connector writes stay under the connector workspace.

• Source folders are not modified by pointer registration.

• _AI_CHATS_ОБЩИЕ is not reorganized by this app.

• Large outputs are written to results/ and read by slice.

• Calls are logged by hash, not full arguments.

• Widget resource uses text/html;profile=mcp-app.

V6 additions

V6 adds a verified 100-improvement registry and audit layer:

• list_implemented_improvements

• audit_100_improvements

• create_support_bundle

• runtime_metrics

• validate_connector_config

• create_debug_snapshot

Files:

• docs/V6_100_IMPLEMENTED_IMPROVEMENTS.md

• docs/V6_100_IMPLEMENTED_IMPROVEMENTS.json

Test:

npm run test:v6

V7 no-window startup and 1000 forward improvements

Manual npm start no longer keeps a visible server console open. It starts the connector in the background through PowerShell 7.

Best daily start method with no popup:

wscript.exe .\CompanionConnector-START-HIDDEN.vbs

Or double-click:

CompanionConnector-START-HIDDEN.vbs

Install hidden startup at Windows logon:

.\scripts\install-hidden-task.ps1

Remove hidden startup task:

.\scripts\uninstall-hidden-task.ps1

V7 also adds a 1000-item forward improvement backlog. It is explicitly not falsely marked as implemented:

• docs/V7_1000_FORWARD_IMPROVEMENTS.md

• docs/V7_1000_FORWARD_IMPROVEMENTS.json

V7 tests:

npm run test:v7
npm run test:windowless

V8 read-only full-folder bridge

V8 adds tools for large read-only folder access without modifying the source folder:

• folder_explorer

• create_readonly_folder_manifest

• audit_readonly_folder_manifest

• create_readonly_folder_content_bundle

• read_folder_bundle_chunk

• search_folder_content_bundle

• create_fable_folder_handoff

Tested target:

J:\ПРОЕКТЫ\G01_All_About_Trading\G01_P09_All_for_TradingView\G01_P09_01_Project\TradingView_INDICATORS\IGOR_ENTER2\research_out

V8 test:

npm run test:v8

The source folder is only read. Manifest, chunks, hashes, and Fable guide are written under CompanionConnector results/.

V14 documents, archives, and web snapshots

V14 adds document/archive/web research tools:

• document_toolchain_report

• inspect_document_file

• inspect_archive_file

• extract_archive_to_results

• create_web_snapshot

• universal_resource_inspect

Supported document families include text/HTML/CSV/JSON/code files, DOCX, PPTX, XLSX/XLSM, and PDF when a Python PDF engine is available.

Archives are inspected read-only and extracted only into connector results/ with path-traversal protection.

Test:

npm run test:v14

V15 project research map

V15 adds project-level intake and relationship mapping:

• create_folder_research_map

• inspect_linked_resources_from_file

• create_project_intake_bundle

These tools create Markdown/JSON outputs with file classification, extension counts, URL links, local references, import/include/source-like lines, and research intake summaries.

Test:

npm run test:v15

V16 browser, live visual monitor, password archives, advanced charts, stronger Fable model

V16 adds browser/DOM tools, repeated screenshot capture, password archive support, advanced chart image analysis, local model controls, and durable full Fable micro-read controls.

New V16 tools include:

browser_start
browser_list_tabs
browser_new_tab
browser_navigate
browser_dom_snapshot
browser_screenshot
desktop_screenshot
browser_click_text
browser_type_selector
browser_press_key
browser_live_monitor
inspect_password_archive
extract_password_archive_to_results
analyze_chart_advanced
local_model_status
local_model_pull
local_model_chat_test
start_full_fable_micro_read
get_full_fable_micro_status

V17/V18 human live, semantic index and queue

V17 adds human-style desktop window listing, active-window inspection, focus, desktop screenshots, screen OCR, live watch screenshots, mouse/keyboard actions, and lightweight semantic indexing/search. V18 adds durable queue jobs, queue health, queue execution, cancel, and path safety audit. See docs/V17_V18_HUMAN_LIVE_SEMANTIC_QUEUE_STATUS.md.

V19 Live Agent Loop

V19 adds a real observe-plan-act loop:

• live_agent_observe captures screen, OCR, active window, windows list, and optional browser state.

• live_agent_fable_plan asks Fable for a JSON action plan from the observation.

• live_agent_apply_action applies one safe allowlisted action, dry-run unless execute=true.

• live_agent_cycle performs observation, Fable planning, optional action, and an after-snapshot log.

All live-agent cycles write evidence into results/<sessionId>/.

V20 Fable Authority Mode

V20 adds an auditable authority layer so Fable5 decisions are recorded instead of only reported verbally.

New tools:

• fable_authority_proposal

• fable_authority_disagreement

• fable_authority_decision_log

• fable_authority_dashboard

• fable_autopilot_dry_run

• fable_autopilot_execute

Authority outputs:

results\fable_authority\decision_log.jsonl
results\fable_authority\dashboard.json
results\fable_authority\dashboard.html

Live-agent proof chain now records observation, Fable plan, executed/dry-run action, next observation, and dashboard evidence.

V21 Direct Fable5 Inbox

Direct user-to-Fable5 task entry is available through:

.\Fable5-Direct.ps1 "Ask Fable5 something directly"

Dashboard:

.\Fable5-Direct-Dashboard.ps1

MCP tools:

• fable_direct_submit

• fable_direct_inbox

• fable_direct_read

• fable_direct_dashboard

All direct messages are logged under results\fable_direct and mirrored into results\fable_authority\decision_log.jsonl.

V22 Fable5 Direct Mode

New direct route:

• fable5

• fable5_direct_mode_manifest

• fable_capability_snapshot

• fable_capability_review

• fable5_request_chatgpt_help

Trigger words for a connected ChatGPT app:

FABLE5:
@Fable5
F5:

The MCP connection cannot technically replace ChatGPT's base model by itself. The connected app must route these trigger messages to the fable5 tool. The manifest is exposed at companion://fable5-direct-mode and via fable5_direct_mode_manifest.

V23 direct CLI fix

The local direct command now accepts natural multi-argument PowerShell input and preserves UTF-8 Cyrillic text:

.\Fable5-Direct.ps1 Yeah. "Запусти новый Хром."
.\Fable5-Direct.ps1 "Запусти новый Хром."
.\Fable5-Direct.ps1 -Strong "сложная задача для Fable5"

-Strong routes to local qwen2.5:3b; default remains the faster Fable path so existing MCP connections do not stall.

See: docs/V23_DIRECT_CLI_FIX_AND_MODEL_STATUS.md.

V24 Fable5 direct execution

V24 changes direct Fable5 mode from reply-only into safe execution mode.

Direct CLI:

.\Fable5-Direct.ps1 FABLE5: "Запусти новый Хром"

The MCP tool fable5 now routes through the direct executor and writes proof logs under:

results\fable_direct_exec\

New tool:

• fable5_execute

Test:

npm run test:v24

Important: the executor can read visible screen/OCR and controlled browser DOM where available. Full hidden ChatGPT history requires transcript export/pass-through or browser DOM access.

V25 ChatGPT chat reader executor

V25 adds a direct execution path for tasks like:

FABLE5: прочти чат https://chatgpt.com/c/... и выведи русское сообщение по пунктам

The fable5_execute tool now detects ChatGPT chat URLs, opens the URL through the default user browser, searches for likely windows such as Codex Chat Watch or ChatGPT Classic, captures repeated OCR, summarizes the visible result in Russian, displays the result on screen, and writes a proof log to results/fable_direct_exec.

Test:

npm run test:v25

V26: Context + UIA + visible output

V26 improves FABLE5: direct execution for ChatGPT chat-reading tasks.

Main behavior:

• ChatGPT conversation URLs are recognized.

• ChatGPT Classic / browser windows are read with UI Automation and OCR where possible.

• The current chat context can be passed into the tool, so FABLE5: прочти прошлое сообщение can work when ChatGPT routes the previous message as context.

• Results are shown visibly in Notepad and an HTML page, not only saved in JSON.

• Every run still writes a proof log under results/fable_direct_exec.

Use:

.\Fable5-Direct.ps1 "FABLE5: прочти чат https://chatgpt.com/c/... и дай вывод по-русски"

For the ChatGPT/MCP app, route FABLE5: messages to fable5_execute and include recent chat text in context when the request asks to read the current or previous message.