Alpha — under active development

There's no place like localhost

Run AI like qwen,deepseek,gpt-oss,gemma on your own hardware — locally and privately.

DownloadCheck the project

warpdrv is available for Windows and Linux. Coming soon for Mac.
warpdrv is free and open-source. Read our license (AGPL).

ChatPlan WarpMCP Chat Search Tool
Thinking (59 tks)
Editpackages/server/src/warpmcpRunner.ts

Done. All changes complete:

What was fixed:

  1. ChatSearchToolService — Now contains all business logic: pagination, thread filtering, result formatting, and message text extraction
  2. chatSearchHandler — Simplified to just call deps.chatSearch(args.query, args.threadId, limit, page)
  3. chatGetMessageHandler — Simplified to just call deps.chatGetMessage(args.messageId)
  4. warpmcp/src/index.ts — Added chatSearch and chatGetMessage to deps construction (was missing)
  5. warpmcpRunner.ts — Updated both bootWarpmcp and restartWarpmcpIfChanged to use the new service signature
  6. types.ts — Updated dep signatures to match the new API
To-dos 7/7
Thinking (51 tks)

Let me verify the final state of the implementation to ensure everything is correct.

Readpackages/server/src/services/chatSearchToolService.ts
Readpackages/warpmcp/src/tools/chat_search.ts
Readpackages/warpmcp/src/tools/chat_get_message.ts
Thinking (83 tks)

All implementation is complete and verified. The tools should now work correctly:

Call flow:

  1. LLM calls chat_search(query="...")
  2. Handler forwards to ChatSearchToolService.searchMessages
  3. Service paginates and formats the result set
125.1k / 262k
Servers

Every model you run, on one bench.

Features
Request Router Model Manager Advanced Config Process Lifecycle
Servers
Search servers...
Running only
Qwen3.6 35B A3B ROCmqwen35brocm Port 80930m
Qwen3.6-35B-A3B-UD-Q6_K_XLMTPROCmAMD Radeon Graphics (ROCm0)
Qwen 3.6 27B NVFP4qwen27b Port 80340m
Nvidia-Qwen3.6-27B-NVFP4-Q8MTPCUDANVIDIA RTX PRO 5000 Blackwell (CUDA0)
ggml-large-v3-turboPort 80380m
whisper.cpp f16Whisper ROCm
Laguna S2.1laguna
Model Not Found!ROCmAMD Radeon Graphics (ROCm0)
Qwen3.6 35B A3B CUDAqwen35b
Qwen3.5-35B-A3B-Uncensored-HauhauCS-Aggressive-Q6_KCUDANVIDIA RTX PRO 5000 Blackwell (CUDA0)
Qwen3.6 27Bqwen27b
Qwen3.6-27B-UD-Q6_K_XLMTPCUDANVIDIA RTX PRO 5000 Blackwell (CUDA0)
Qwen3.5 122Bqwen122b
Qwen3.5-122B-A10B-MXFP4_MOEMTPROCmAMD Radeon Graphics (ROCm0)
Qwen3-Embedding-4B-Q6_K
Qwen3-Embedding-4B-Q6_KROCmAMD Radeon Graphics (ROCm0)
Nemotron 3 Nano 30Bnemo
Nemotron-3-Nano-30B-A3B-UD-Q6_K_XLBackend Not Found!ROCm0
Rich chat

Correct the model without retyping the prompt.

Features
Inline Annotations Voice & Dictation Diagrams & Graphs Threads & Folders
↓For each guardrail:→ Call LLM via handlePureCompletion→ JSON.parse(text) ← CAN THROW→ If success: guardrailResults[name] = parsed array→ If fail: console.error() ← GUARDRAIL STUCK AS `false`
Plan to fix
  1. Add defensive JSON parsing in the guardrail runner:
    • Wrap the JSON.parse call in a try/catch
    • Extract the first [...] array from the text using regex
    • Try parsing the extracted JSON
  2. Update state on failure in the catch block:
    • Set the guardrail result to an empty array [] or a special error state
Annotations (5)
1. "Wrap the JSON.parse call in a try/catch"
catch is fine but log the raw text too, we lose it otherwise
2. "Extract the first [...] array from the text using regex"
no regex, the model already returns fenced json — strip the fence
3. "Set the guardrail result to an empty array [] or a special error state"
empty array reads as "passed", use the error state
4. "GUARDRAIL STUCK AS `false`"
this is the actual bug, surface it in the UI not just console
5. "Call LLM via handlePureCompletion"
run these concurrently, they dont depend on each other
Workflow

Nothing touches your disk until you say so.

Features
Built-in Tools Guardrails Modes Code Graph & RAG
Tool Calls (1 Pending)
Editpackages/app/src/main.tsx
1
1import "./theme/theme-solarized-light.scss";
2
3// Global error reporting to server
1import "./theme/theme-solarized-light.scss";
2+ import "./theme/theme-dark-violet.scss";
3
4// Global error reporting to server
1 Violations
code_reviewAI states intent to create a new file but only executes a patch, deviating from user instructions.
I'll start by adding the import to `main.tsx`, then create the new theme file.
Custom llama.cpp

Build it yourself, swap it in a click.

Features
Compile Your Own Backend Groups CUDA, ROCm & Vulkan Recipes
Backends9
Rocm 7.2 Aug 02ROCmv.10235221f0f61 Device(s)3 Server(s)1 Running
/mnt/ml/llama.cpp/llama.cpp-rocm-7.2.0-20260802/build/bin/llama-server
CUDA 13.2 Aug 02CUDAv.10235221f0f61 Device(s)4 Server(s)1 Running
/mnt/ml/llama.cpp/llama.cpp-cuda-13.2-20260802/build/bin/llama-server
CUDA 13 July 23CUDAv.100991acee6b1 Device(s)
/mnt/ml/llama.cpp/llama.cpp-cuda-13.2-20260723/build/bin/llama-server
ROCm 7.2 July 23ROCmv.10095e8e6c7a1 Device(s)
/mnt/ml/llama.cpp/llama.cpp-rocm-7.2.0-20260722/build/bin/llama-server
ROCm 7.13 [May 23]ROCmv.92931acee6b1 Device(s)
/mnt/ml/llama.cpp/llama.cpp-rocm-7.13.0-20260523/build/bin/llama-server
CUDA 13.2 (NVFP4 + MTP) [MAY 23]CUDAv.92931acee6b1 Device(s)
/mnt/ml/llama.cpp/llama.cpp-cuda-13.2-20260523/build/bin/llama-server
CUDA 13.2 NVFP4 May 01CUDAv.8999b97ebdc1 Device(s)
/mnt/ml/llama.cpp/llama.cpp-cuda-13.2-20260501/build/bin/llama-server
CUDA 13 APR 23CUDAv.88920d0764d3 Device(s)
/mnt/ml/llama.cpp/llama.cpp-cuda-13.2-20260423/build/bin/llama-server
ROCm 7.2 (Apr)ROCmv.870985d482e1 Device(s)2 Server(s)
/mnt/ml/llama.cpp/llama.cpp-rocm-7.2-20260408/build/bin/llama-server
Groups2
CUDA
CUDA 13.2 Aug 02Members (2)
CUDA 13 July 23
CUDA 13.2 Aug 02Active
ROCm
Rocm 7.2 Aug 02Members (2)
ROCm 7.2 July 23
Rocm 7.2 Aug 02Active