feat: add CLI command to download local-inference models (no goosed server required)

[eugenio]

Problem

Downloading local GGUF models for the local-inference feature currently requires:

1. Starting the goosed agent HTTP server
2. Knowing the secret key (randomly generated, printed to stderr as GOOSE_SERVER__SECRET_KEY or needing GOOSE_SERVER__SECRET_KEY env var)
3. Using curl with the correct API endpoint, headers, and URL-encoded model IDs:

   curl -sk -X POST -H "X-Secret-Key: $SECRET" -H "Content-Type: application/json" \
     -d '{"spec":"bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M"}' \
     https://127.0.0.1:3000/local-inference/download

4. Monitoring progress via a separate GET endpoint
5. Stopping the goosed server when done

This is especially painful when:

• Configuring local provider via goose configure — it says "Model not found" with no hint about downloading
• Setting up the toolshim interpreter — GOOSE_TOOLSHIM=true fails with cryptic Ollama errors if no local model is downloaded
• Running headless / on servers where there is no desktop UI

Proposed Solution

Add a CLI subcommand for model management:

# List available models and their download status
goose models list

# Download a model
goose models download bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M

# Show download progress
goose models status

# Remove a downloaded model
goose models remove <model-id>

Benefits

• No need to start goosed server just to download a model
• goose configure for Local Inference could auto-prompt to download if model not found
• Toolshim setup becomes: set env var + goose models download — no Ollama, no goosed

Current Workaround

Start goosed with a known secret key and use the HTTP API:

$env:GOOSE_SERVER__SECRET_KEY = "localtest123"
goosed agent
# In another terminal:
curl -sk -X POST -H "X-Secret-Key: localtest123" -H "Content-Type: application/json" -d '{"spec":"bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M"}' https://127.0.0.1:3000/local-inference/download

Related

• feat: allow configuring provider for developer extension's message augmentation (hardcoded Ollama dependency) #8270 — Developer extension hardcodes Ollama for message augmentation
• bug: Toolshim OllamaInterpreter returns noop for all streaming tool call fragments #8275 — Toolshim streaming fragment handling

[vincenzopalazzo]

Hey! The use case you're describing is already covered by the existing goose local-models CLI subcommand (alias goose lm). It's enabled by default and doesn't require the goosed server at all.

Here's how the current commands map to what you proposed:

goose local-models list                          # list downloaded models and status
goose local-models download <user/repo:quant>    # download a GGUF model directly
goose local-models search <query>                # search HuggingFace for GGUF models
goose local-models delete <id>                   # remove a downloaded model

The download command already handles progress reporting inline (polls every 500ms and shows percentage + MB downloaded).

For the status command — download progress is currently shown as part of the download command itself, so there's no separate status check. If you need to check whether a model is already downloaded, goose local-models list shows a check/cross indicator for each model.

The shorter aliases also work: goose lm list, goose lm download, etc.

If you think a goose models alias (in addition to local-models and lm) would improve discoverability, that's a small change we could make. But the core functionality is already there.

[eugenio]

Thanks, I hadn't figured it out... Kudos for the suggestion! Il giorno ven 3 apr 2026 alle ore 13:16 Vincenzo Palazzo < ***@***.***> ha scritto:

…

*vincenzopalazzo* left a comment (block/goose#8282) <#8282 (comment)> Hey! The use case you're describing is already covered by the existing goose local-models CLI subcommand (alias goose lm). It's enabled by default and doesn't require the goosed server at all. Here's how the current commands map to what you proposed: goose local-models list # list downloaded models and status goose local-models download <user/repo:quant> # download a GGUF model directly goose local-models search <query> # search HuggingFace for GGUF models goose local-models delete <id> # remove a downloaded model The download command already handles progress reporting inline (polls every 500ms and shows percentage + MB downloaded). For the status command — download progress is currently shown as part of the download command itself, so there's no separate status check. If you need to check whether a model is already downloaded, goose local-models list shows a check/cross indicator for each model. The shorter aliases also work: goose lm list, goose lm download, etc. If you think a goose models alias (in addition to local-models and lm) would improve discoverability, that's a small change we could make. But the core functionality is already there. — Reply to this email directly, view it on GitHub <#8282?email_source=notifications&email_token=AACHMZAODZGAL2E7LWYOZ6L4T6MQTA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIMJYGMYDKMJYGIYKM4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2KYZTPN52GK4S7MNWGSY3L#issuecomment-4183051820>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AACHMZEQEENL3HGSOSH3AFT4T6MQTAVCNFSM6AAAAACXLVGSE6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DCOBTGA2TCOBSGA> . You are receiving this because you authored the thread.Message ID: ***@***.***>