What is mcp-ollama (~alter)?
mcp-ollama, published as @truealter/mcp-ollama by Alter Meridian, the Australian company behind the ~alter identity service, lets a cloud agent such as Claude Code pass routine work to a model running in Ollama on the same machine. Its ten tools are shaped around jobs rather than API calls: summarise, classify or tag, draft to a convention, write docstrings and tests, review code, turn a diff into a commit message or changelog, convert formats, and have a vision model describe a screenshot.
It is very new and very small: created on 1 September 2026, Apache-2.0, TypeScript, 1 GitHub star and about 75 npm downloads a week. The latest npm release is 0.2.1 from 3 September 2026, and it is listed in the official MCP registry.
How to install mcp-ollama (~alter)
npx -y @truealter/mcp-ollamaIs mcp-ollama (~alter) worth it?
mcp-ollama answers a different question from ollama-mcp. It does not manage Ollama; it gives a cloud agent a cheaper pair of hands. The ten tools carry their own prompts, so local_diff with commit-message asks for imperative mood and a subject under 72 characters, and local_code offers docstring, test, explain, review, types, comments and refactor-suggest. Model management is limited to local_models and local_pull. There is no delete, copy, create or push tool, so an agent cannot remove or publish your models through it.
The security choices are the best part, in our reading of the source. It speaks stdio only and opens no port. Ollama's local API on port 11434 needs no authentication, so the server refuses at startup any OLLAMA_HOST other than localhost, 127.0.0.1 or ::1 unless MCP_OLLAMA_ALLOW_REMOTE=1 is set, which stops a changed setting quietly sending prompts elsewhere. Model names are checked against a pattern that rejects a registry host, because Ollama treats a dotted first segment as an alternative registry and would download from it. local_vision reads image files from paths the caller supplies, but only PNG or JPEG, capped at 32 MB each.
The defaults are tuned for a single GPU. The text model is hermes3:8b and the vision model qwen2.5vl:7b, with vision context capped at 8,192 tokens and the model held for 30 minutes. Output ceilings are fixed in code (512 tokens for a commit message, 4,096 for code and transforms), and the README warns that the vision model invents data below about 1,280 pixels wide. There is no streaming, queue or cache, and no Ollama Cloud support.
The catch is everything around the code. The README is partly out of date: it says local_pull rejects tagged names, yet the published 0.2.1 accepts them. Much of the README is a pitch for ~alter's identity product, and if the ~alter runtime is installed the server appends token counts to a file in your home directory. Nothing leaves the machine, but it is worth knowing. With one maintainer company and almost no users, the risk is abandonment rather than bad engineering.
Where it earns its place
A developer on a metered Claude Code plan
Write commit messages and PR descriptions without spending frontier-model tokens
The agent passes git diff --staged to local_diff with commit-message or pr-description, and hermes3:8b drafts it on the local GPU. The cloud model only reads the short result.
A front-end engineer checking UI states
Find out whether a screenshot shows real data, an empty state or an error
local_vision sends the PNG to qwen2.5vl:7b and returns what is on screen and any clipped text. Feed it near full resolution; it does not rank severity or approve the design.
A support lead at a 20-person software company
Tag a batch of exported tickets by product area without sending them to a cloud API
The agent loops over the file and calls local_analyze with a fixed list of tags; customer text stays on the machine and only the labels go back into the agent's context.
Before you commit
- Setup
- ollama pull hermes3:8b, then claude mcp add --transport stdio ollama -- npx -y @truealter/mcp-ollama. Needs Node 18 or newer. Use the scoped name; the unscoped mcp-ollama on npm belongs to another publisher. Docker needs a local build and MCP_OLLAMA_ALLOW_REMOTE=1.
- Integrations
- 10 tools: local_generate, local_summarize, local_analyze, local_draft, local_code, local_diff, local_transform, local_models, local_pull and local_vision. Only local_pull changes anything on disk.
- Data and privacy
- Stdio only, no port, no telemetry. Calls only OLLAMA_HOST, loopback by default. local_vision can read any PNG or JPEG path it is given, up to 32 MB. If the ~alter runtime is installed, each generation appends tool, model and token counts to ~/.local/share/alter-runtime/token-burn.jsonl. No published advisories; reports go to security@truealter.com.
- Cost at scale
- Free; the cost is GPU time. The README suggests qwen2.5-coder:32b for code review, which needs a large card, and a 14B model for prose on a 16 GB card.
- Limits
- No streaming, queueing or caching. max_tokens only on local_generate (2,048 default) and local_summarize (1,024). Vision context capped at 8,192 tokens. Only the latest 0.x release gets security fixes.
What’s good
- Refuses non-loopback Ollama hosts by default
- No tools that delete, copy or publish models
- Task-shaped tools with built-in prompts
- Local vision tool with file type and size checks
What’s not
- One release, 1 star and about 75 weekly downloads
- README out of step with the published code
- Fixed output ceilings on most tools
- No Ollama Cloud, streaming or model removal
Best for
Developers paying per token for a cloud agent who already run Ollama on a decent GPU and want routine text chores kept local.
mcp-ollama (~alter) alternatives
mcp-ollama (~alter): common questions
- Is mcp-ollama (~alter) worth it?
- Careful, narrow code for offloading cheap chores to a local model, held back mainly by being a two-month-old project with one star. We rate it 3.5 out of 5.
- How much does mcp-ollama (~alter) cost?
- Free and open source (Apache-2.0). Runs only against your own Ollama, so the cost is your hardware.
- What are the downsides of mcp-ollama (~alter)?
- One release, 1 star and about 75 weekly downloads; rEADME out of step with the published code; fixed output ceilings on most tools; no Ollama Cloud, streaming or model removal.
- Who is mcp-ollama (~alter) best for?
- Developers paying per token for a cloud agent who already run Ollama on a decent GPU and want routine text chores kept local.
- How much does mcp-ollama (~alter) cost for a team or in production?
- Free; the cost is GPU time. The README suggests qwen2.5-coder:32b for code review, which needs a large card, and a 14B model for prose on a 16 GB card.
- Is mcp-ollama (~alter) safe for company data?
- Stdio only, no port, no telemetry. Calls only OLLAMA_HOST, loopback by default. local_vision can read any PNG or JPEG path it is given, up to 32 MB. If the ~alter runtime is installed, each generation appends tool, model and token counts to ~/.local/share/alter-runtime/token-burn.jsonl. No published advisories; reports go to security@truealter.com.
- What are the best mcp-ollama (~alter) alternatives?
- The closest alternatives we have reviewed are Ollama MCP Server.
Answers checked Nov 11, 2026.