Local model hosting · vision sidecar · sandboxed agents

lmcoder Debloated model serving & hardened coding sandboxes.

The primary control hub for local LLMs and multimodal vision sidecars, debloating complex runtimes and routing seamlessly between laptops and dedicated RTX inference servers.

go install ubunatic.com/lmcoder/cmd/lmcoder@latest

Engineering story View source Product home

Project snapshot · measured 2026-10-26

Key facts

Commits 983Last commit 2026-10-26 Source receipt ↗ Commits · 7 days 141 Source receipt ↗ Tickets closed 165 / 225 Source receipt ↗ Studies 36 Source receipt ↗

12-week commit activity

2026-10-26

Aug 3

Aug 10

Aug 17

156

Aug 24

Aug 31

Sep 7

Sep 14

140

Sep 21

Sep 28

Oct 5

Oct 12

Oct 19

View 12-week data table

Week starting	Commits
2026-08-03	0
2026-08-10	0
2026-08-17	2
2026-08-24	156
2026-08-31	23
2026-09-07	1
2026-09-14	69
2026-09-21	140
2026-09-28	86
2026-10-05	96
2026-10-12	77
2026-10-19	63

Key Features

Transparent proxy routing between local laptop endpoints and upstream dedicated RTX inference servers.
Self-deploying and cross-compiling binary installation powered by goself.
Runs coding assistants in rootless Podman containers with single-writable-mount workspaces.
Multimodal Florence-2 vision sidecars with idle timeouts, candidate streaming, and box filtering.
Memory budget validation and automated Multi-Token Prediction (MTP) draft acceleration.

Verified Engineering Milestones

Isolated coding checks and stack install (#047, #156, e96a0a0)
Managed proxy lifecycle and sampling profile injection (#164, 8037408, d5571f3)
Florence-2 vision sidecar, box filtering and streaming preview (#223, #224, ad8d8b7)

Stack

Go 1.25.0, Cobra, YAML specs, llama.cpp, OpenAI-compatible HTTP/SSE, rootless Podman, systemd, Vulkan/ROCm, Florence-2

Why lmcoder

Model hosting tuned to developer hardware

Running coding agents against a local LLM sounds simple until the agent needs shell access, the model blows past GPU memory budgets on shared hardware, or a dependency bump silently breaks streaming. Most setups settle for pointing an agent at an unmonitored endpoint with zero sandboxing.

lmcoder turns personal setup customizations into robust software features: debloating the runtime stack, managing memory limits, accelerating inference with MTP draft heads, and isolating agent execution in rootless sandboxes.

Architecture & Capabilities

Built for dedicated GPUs, laptops & remote servers

Seamless Proxy Routing

Routes agent requests transparently between local laptop inference and an upstream RTX server (such as x600), injecting sampling profiles (fast, default, deep) dynamically.

Self-Installing via goself

Deploys, cross-compiles, and updates itself seamlessly across machines using goself, eliminating remote toolchain dependencies.

Multimodal & Vision Sidecar

Runs Florence-2 and vision embedding sidecars with idle timeouts, candidate preview streaming, and degenerate bounding-box filtering.

Hardened Podman Sandboxes

Executes coding agents inside rootless Podman containers with single-writable workspace mounts and automated zombie reaping.

MTP & Memory Budgets

Manages llama-server instances with automated Multi-Token Prediction draft acceleration and strict VRAM/RAM threshold guards.

CLI Workflow

Unified model management

# Serve a model with memory overload guards and MTP acceleration
$ lmcoder serve Qwen3.8-27B-UD-Q4_K_M.gguf

# Route client traffic upstream to an RTX host with dynamic profile injection
$ lmcoder proxy --backend-host 192.168.1.100 --profile fast

# Run sandboxed agent canary against an isolated workspace
$ go run ./scripts/agent-canaries --agent pi --task coding-hello

Project details

Stack

Go 1.25.0, Cobra, YAML specs, llama.cpp, OpenAI-compatible HTTP/SSE, rootless Podman, systemd, Vulkan/ROCm, Florence-2

Milestones

Isolated coding checks and stack install (#047, #156, e96a0a0) · Managed proxy lifecycle and sampling profile injection (#164, 8037408, d5571f3) · Florence-2 vision sidecar, box filtering and streaming preview (#223, #224, ad8d8b7)

Source

https://codeberg.org/ubunatic/lmcoder ↗

Architecture

Layer composition

Terminal: CLI, proxy controls, service manager
Desktop: Linux, systemd
Inference / AI: llama.cpp, OpenAI-compatible API, Florence-2, MTP, Vulkan/ROCm
Containers / VM: rootless Podman
Web: —

Engineering story

lmcoder: the story behind the project

Second-Vendor Reviews : > - Audited by claude:sonnet:med (session story-reviewer-sonnet ): Approved with corrections . Clarified test counts (19 tests across test_lifecycle.py and test_scan.py ; 14 tests in test_scan.py ), replaced mixed me…

Read the engineering story

Measurement source · 2026-10-26

Verification receipts

Commits: 983 · Last commit 2026-10-26
git -C <repo> rev-list --count HEAD; git -C <repo> rev-list --count --since=<ISO-week-Monday> --until=<ISO-week-Sunday> HEAD; git -C <repo> rev-list --count --since=<measured-date-6-days>T00:00:00 --until=<measured-date>T23:59:59 HEAD; git -C <repo> log -1 --format=%cs HEAD; ls <repo>/issues | grep '^[0-9].*\.md$'; grep -l -E '^\*\*Status\*\*: *(Closed|Done|Resolved|Fixed|Complete)' <repo>/issues/*.md | wc -l; ls <repo>/docs/studies/*.md
repositories behind spec project ids in ~/projects; private-repo filter is a safety net and matches no spec project today; missing repos reported separately; seven calendar days ending on measured date in the measuring host's local timezone; commit counts include all authors and merge commits reachable from HEAD · measured 2026-10-26
Commits · 7 days: 141
git -C <repo> rev-list --count HEAD; git -C <repo> rev-list --count --since=<ISO-week-Monday> --until=<ISO-week-Sunday> HEAD; git -C <repo> rev-list --count --since=<measured-date-6-days>T00:00:00 --until=<measured-date>T23:59:59 HEAD; git -C <repo> log -1 --format=%cs HEAD; ls <repo>/issues | grep '^[0-9].*\.md$'; grep -l -E '^\*\*Status\*\*: *(Closed|Done|Resolved|Fixed|Complete)' <repo>/issues/*.md | wc -l; ls <repo>/docs/studies/*.md
repositories behind spec project ids in ~/projects; private-repo filter is a safety net and matches no spec project today; missing repos reported separately; seven calendar days ending on measured date in the measuring host's local timezone; commit counts include all authors and merge commits reachable from HEAD · measured 2026-10-26
Tickets closed: 165 / 225
git -C <repo> rev-list --count HEAD; git -C <repo> rev-list --count --since=<ISO-week-Monday> --until=<ISO-week-Sunday> HEAD; git -C <repo> rev-list --count --since=<measured-date-6-days>T00:00:00 --until=<measured-date>T23:59:59 HEAD; git -C <repo> log -1 --format=%cs HEAD; ls <repo>/issues | grep '^[0-9].*\.md$'; grep -l -E '^\*\*Status\*\*: *(Closed|Done|Resolved|Fixed|Complete)' <repo>/issues/*.md | wc -l; ls <repo>/docs/studies/*.md
repositories behind spec project ids in ~/projects; private-repo filter is a safety net and matches no spec project today; missing repos reported separately; seven calendar days ending on measured date in the measuring host's local timezone; commit counts include all authors and merge commits reachable from HEAD · measured 2026-10-26
Studies: 36
git -C <repo> rev-list --count HEAD; git -C <repo> rev-list --count --since=<ISO-week-Monday> --until=<ISO-week-Sunday> HEAD; git -C <repo> rev-list --count --since=<measured-date-6-days>T00:00:00 --until=<measured-date>T23:59:59 HEAD; git -C <repo> log -1 --format=%cs HEAD; ls <repo>/issues | grep '^[0-9].*\.md$'; grep -l -E '^\*\*Status\*\*: *(Closed|Done|Resolved|Fixed|Complete)' <repo>/issues/*.md | wc -l; ls <repo>/docs/studies/*.md
repositories behind spec project ids in ~/projects; private-repo filter is a safety net and matches no spec project today; missing repos reported separately; seven calendar days ending on measured date in the measuring host's local timezone; commit counts include all authors and merge commits reachable from HEAD · measured 2026-10-26
12-week commit activity
git rev-list --count --since=<Monday> --until=<Sunday> HEAD
12 latest complete ISO weeks; weekly values are from spec/metrics.yaml.