# Reported issues for lm

Pod holds 6 of 6 problems reported by people outside the maintainer team. Issues filed by the project's own owners, members and collaborators are excluded entirely — a maintainer's release checklist is not a warning to a prospective user.

Back to [lm](/mcp/lm).

## Most discussed

### [Provider] Lower-cost DeepSeek/Kimi/GLM endpoint for testing

Hi,

I noticed Houtini LM and that users care about token cost / provider switching for Claude Code delegation.

I'm testing a lower-cost China-model API (DeepSeek / Kimi / GLM / Qwen) via an Anthropic/OpenAI-compatible endpoint.

Would you be open to an optional cloud backend for testing? I'll provide free test credits and a price/latency comparison vs OpenRouter — no architecture change.

Which backend do your users use most today for cloud delegation?

Thanks.

[Read the thread](https://github.com/houtini-ai/houtini-lm/issues/13) · 2026-06-26 · open · external user · 0 comments

### feat: Add path remapping support for containerized MCP servers (e.g., Podman/Docker)

**Problem:**
When the Houtini-LM MCP server runs inside a container (e.g., via Podman or Docker), it cannot access files on the host filesystem unless:

The host directory is mounted into the container (via -v), and
The file paths passed to code_task_files match exactly the container’s internal path structure.
This creates two practical issues:

🔧 Fragile configuration: Users must ensure host and container paths align (e.g., /home/user/project → /home/user/project), which is error-prone if mount

[Read the thread](https://github.com/houtini-ai/houtini-lm/issues/12) · 2026-06-25 · open · external user · 0 comments

### Auto-router sends requests to deleted model, ignores HOUTINI_LM_MODEL env var

## Describe the bug

After deleting a model from Ollama and pulling a replacement,
the auto-router continues to send requests to the deleted model (404 error).
Neither `HOUTINI_LM_MODEL` env var nor cache cleanup resolves it.

## Steps to reproduce

1. `ollama pull qwen2.5:7b`
2. Use houtini-lm (auto-router learns qwen2.5:7b)
3. `ollama rm qwen2.5:7b`
4. `ollama pull qwen2.5-coder:7b`
5. `ollama list` — shows only qwen2.5-coder:7b
6. `discover` — shows active model: qwen2.5-coder:7b ✅
7. Call `c

[Read the thread](https://github.com/houtini-ai/houtini-lm/issues/9) · 2026-05-06 · open · external user · 0 comments

### Automatically adjust token count and timeout

1. code_task uses a static max_tokens value regardless of input size or model context length. On small models (4K–8K context), a large input can leave almost no room for output, producing truncated or empty analysis. On large-context models, the budget is needlessly conservative.
2. The MCP SDK imposes a hard ~60s timeout on tool calls. Three failure modes currently exist: 
 - Slow TTFT on large inputs can exhaust the timeout before generation even starts
 - A stalled mid-stream connection hangs

[Read the thread](https://github.com/houtini-ai/houtini-lm/issues/5) · 2026-03-30 · closed · external user · 1 comment

### Add support for sending file paths to LLM

The existing code_task tool requires the caller to read files and pass raw content as a string parameter. For anything beyond a single small file, this bloats the MCP client's context window unnecessarily and makes multi-file analysis awkward to orchestrate. 

Suggested solution:
Add a code_task_files companion tool that accepts an array of absolute file paths. The server reads the files itself using Promise.allSettled() (so failures are reported gracefully rather than crashing), concatenates th

[Read the thread](https://github.com/houtini-ai/houtini-lm/issues/4) · 2026-03-30 · closed · external user · 1 comment

### Feature proposal: Automated quality checks for remote LLM completions (optional Claude review)

Dear Richard,

Many thanks for your brilliant work on Houtini – i think it's brilliant.

Since i work mainly on my laptop (m1 pro) i don't have access to a powerful gpu so i would need to use remote APIs for my work.

As delegation to paid remote APIs (e.g., Inception Labs/OpenRouter) becomes more common, would you consider adding an optional feature for Claude (or the orchestrator) to automatically check delegated completions for quality/accuracy before passing them on?

This could help efficie

[Read the thread](https://github.com/houtini-ai/houtini-lm/issues/3) · 2026-03-03 · closed · external user · 2 comments

## Most recent

The remaining reports are on [the project's issue tracker](https://github.com/houtini-ai/lm/issues).