RSCOT · CODE AGENT
How to use

Rescot code agent

Rescot is a local-first code agent that drives real tools on your machine — files, shell, Python REPL, Git, web, and (via ModelRein) direct model control. This page explains what it does, how to use it, and how to connect ModelRein in one of three ways.

Contents

  1. Install & launch
  2. Your first session
  3. Tool catalog · 52 skills
  4. Permissions · what the agent can do
  5. ModelRein · adding model control
  6. Mode 1 · your GPU server over SSH
  7. Mode 2 · GPU on this computer
  8. Mode 3 · an ModelRein address you already have
  9. Doctor-first — probe before you plan
  10. Troubleshooting

1Install & launch

Install Rescot from learningfind.com or the Microsoft Store (listed as Rescot). On first launch it creates a workspace and asks for the API key of the model provider you choose, or a local model.

2Your first session

  1. Press + in the sidebar to start a new session.
  2. Click the folder icon in the Composer to pick a workspace (any folder on your disk).
  3. Type a task. Examples that work out of the box:
    • "Clean up the LaTeX file paper.tex for publication — fix whitespace, section titles, table caption, bibliography."
    • "There's a failing pytest in this repo. Find the bug and fix it with the smallest possible change."
    • "Build a react-vite + tailwind landing page for my project in ./site/."
  4. Watch the agent plan, call tools, show you results. You approve destructive calls per your permission settings.
Rescot sessions are stateful — the agent remembers earlier turns in the current session but starts fresh in a new one. Memory (long-term facts about you and your projects) is separate; see Settings → Memory.

3Tool catalog · 52 skills

The agent has 52 tools grouped into 6 families:

Files · read_file · write_file · edit_file · apply_diff · list_dir · glob · grep · format_file
Shell · run_command · run_powershell · run_process · start_background_task
Python · run_python_repl · verify_package · verify_package_batch
Git · git_diff_summary · read_pr_comments · reply_pr_comment · create_pr
Web · web_search · read_pdf · download_dataset
Environment · check_capability · system_info · gpu_info · list_processes · allocate_port · kill_process · mount_external_folder
UI driver (Windows) · ui_focus_state · ui_click · ui_type · ui_screenshot · ui_window_bounds · ui_send_keys
ModelRein (when enabled) · lmcp_introspect · lmcp_generate · lmcp_compute_vector_auto · lmcp_head_mask · lmcp_layer_route · lmcp_pipeline · … (24 total)

The agent decides which tool to use based on your task. You don't call tools directly.

4Permissions · what the agent can do

Settings → Permissions. Five categories, each with three modes: auto (silent) · ask (modal per call) · deny (blocked).

Access scope is set separately: workspace-only (default; agent sees only the folder you picked), mounted (plus other folders you explicitly mount), or full disk (anywhere; your call).

5ModelRein · adding model control

By default Rescot talks to the model provider you configured for generation, a normal chat completion loop. Connecting ModelRein (formerly LMCP) adds 24 lmcp_ tools in Task mode that reach inside an open-source model: residual steering, attention-head masking, layer routing and cross-model pipelines. There is no pip package or public repository; the ModelRein service runs on a GPU server you control, and Rescot connects to it in one of three ways. Step-by-step: how to use ModelRein.

┌───────────────┐ SSH tunnel ┌──────────────┐ nnsight ┌────────────┐ │ Rescot (your │───────────────▶│ ModelRein service │──────────▶│ open- │ │ desktop app) │ local 8030–8035│ server :8030 │ │ source │ └───────────────┘ └──────────────┘ │ model │ Settings › Connections › LMCP connections └────────────┘

Mode 1 · your GPU server over SSH most common

Rescot → Settings › Connections › LMCP connections → Connect to server. Paste the SSH command your cloud gives you (for example ssh -p 37436 root@connect.example.com) and the password, or leave the password empty to use your ~/.ssh keys. Rescot's built-in SSH client checks the GPU and the ModelRein service, then forwards the server port to the first free local port in 8030–8035. The password stays in memory only.

If the service is not running, use Start LMCP on the server → Run on server. The default command starts python3 -m uvicorn server:app --host 127.0.0.1 --port 8030 from ~/lmcp and logs to ~/lmcp.log; edit it if your server keeps ModelRein elsewhere.

Mode 2 · GPU on this computer best latency

The Local card detects NVIDIA GPUs on this computer. If it shows that the service is not running, start the ModelRein service in its directory with python -m uvicorn server:app --host 127.0.0.1 --port 8030 and press Recheck; Rescot connects on its own.

Mode 3 · an ModelRein address you already have

In the composer's compute chip choose Add address, enter the LMCP server address and, if your server requires one, an access token, then press Connect and use.

Security: ModelRein has no built-in authentication in v0.4. Prefer the SSH tunnel; if you expose the port yourself, put it behind an authenticating proxy or a VPN.

6Doctor-first — probe before you plan

Agent sessions start with environment probes when they need a system capability. If you ask the agent to "generate Chinese narration for this video", it will first call check_capability("tts_zh_cn") to decide whether to use SAPI, Edge TTS, or recommend an install — rather than trying one path and failing five times.

Supported probes: tts_zh_cn · tts_en · ffmpeg · pdflatex · office_word_com · cuda_gpu · python_3_10+ · node · docker · git. The agent chooses to probe; you don't have to prompt for it.

7Troubleshooting

The ModelRein server shows as not connected

For an SSH server, press Reconnect and enter the password again: it is never saved, so a restart of Rescot drops the tunnel. If the tunnel is up but the service does not respond, start it with Run on server and read the output.

Agent is slow on Windows PowerShell operations

The old run_command "powershell ..." path went through cmd.exe and ate quotes. If you see the agent retry PowerShell scripts, make sure you're on Rescot 1.8.73+ which added run_powershell (Base64 UTF-16LE → -EncodedCommand). Blog post.

Agent asks before every file edit — can I speed it up?

Settings → Permissions → writes → auto. Still prompts for deletes.

ModelRein tools are not used

Settings › Connections › LMCP connections → make sure the server is connected and Used in Task mode is on. The tools are offered in Task mode.

Nothing works and I just want to try the agent without any setup

Pick a workspace (any folder, even an empty one), give the agent a task — e.g. "write me a Python script that renames all .jpg files in this folder to a date-prefixed scheme". The agent runs on your configured provider and doesn't need ModelRein for most everyday tasks.

See also

Rescot code agent docs · home · blog

Installing Rescot, pricing and data flow: learningfind.com · install guide