Hyperbench
OfficialClick on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@HyperbenchTrace this sensor's signal from connector to driver and cite code and schematic pages."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Hyperbench Workspace
Local hardware context for the agent you already use.
Connect repositories, PDFs, and a system diagram. Your agent can follow source and document evidence, investigate failures, and review firmware changes through MCP and three small skills.
Open source. Local storage. No code indexer, model keys, account, or hosted backend.
Install
Requires Node.js 22.13+ (a current LTS release is recommended). Git is needed to clone repositories; existing folders and documents can be used without it.
npm install -g https://github.com/metryxa-studio/hyperbench-workspace/releases/download/v0.1.0/hyperbench-workspace-0.1.0.tgz
hyperbench startThis opens the local browser workspace. Keep the terminal process running; Ctrl+C stops it. It does not stop or alter your agent. There is no desktop installer or native compilation step.
To try it without a global install:
npx --yes --package=https://github.com/metryxa-studio/hyperbench-workspace/releases/download/v0.1.0/hyperbench-workspace-0.1.0.tgz hyperbench startAgent configurations refer to the installed package path. Prefer the global install for a durable connection; regenerate your connection if you move or remove the package.
Related MCP server: rag
First project
Create a project.
In Repositories, register a local folder or clone a Git URL to a new local destination. Git uses your existing SSH/credential setup.
In Documents, drop PDFs or text files into the workspace, or import by local path.
In System, add components, ports, and connections. Link them to source paths or document pages. Click Save to keep edits.
Open Connect agent and copy the configuration for your client. Reconnect the client, then ask it to list Hyperbench projects.
The same project can have multiple repositories and documents. An agent can submit a system draft; it appears in the UI for review. Acceptance checks its original revision so a stale draft cannot silently replace newer work.
Example requests:
“Trace this sensor's signal from the connector to its driver and consumers. Cite the code and schematic pages.”
“Investigate why the readings wrap around after this firmware update.”
“Review this PR's packet parser change and run a bounded fuzz test using the existing host-side target.”
Skills
All skill prose is original and provider-independent. They run inside your chosen agent; Hyperbench does not run a model or shell on its behalf.
Skill | Purpose |
Follow code, configuration, schematic, and system relationships with citations. | |
Reproduce a symptom, distinguish causes, and validate a scoped fix. | |
Trace a local PR diff and select focused regression, property, or fuzz checks. |
Copy skills into a directory supported by your agent, for example a project's .agents/skills directory:
hyperbench skills --destination /path/to/project/.agents/skillsExisting skill folders are never overwritten. For clients that do not discover SKILL.md, load the chosen file as instructions. A skill can also be downloaded directly from this repository; its host still needs the MCP connection for Hyperbench tools.
PR review is invoked by the user or their existing agent workflow. This package does not watch GitHub, fetch private PRs automatically, or send review comments.
Agent connection
Use the UI's project-specific configuration, or print one:
hyperbench connect codex --project YOUR_PROJECT_ID
hyperbench connect claude --project YOUR_PROJECT_ID
hyperbench connect generic --project YOUR_PROJECT_IDThese commands print settings; they do not overwrite existing client configuration. See agent setup for supported formats and troubleshooting.
The MCP server uses stdio, with the same local data directory as the browser workspace. It can run without the browser process. --project restricts tools to one project; omitting it makes all local workspace projects available.
Tools | Access |
| Available projects and their local context |
| Bounded file access and literal search, without an index |
| Extracted text and rendered PDF pages/crops |
| Saved architecture and pending proposals |
| Create a draft; cannot accept it or overwrite the saved model |
Your agent retains its own model, API keys, script execution, and permission controls. Hyperbench has no provider routing or agent harness.
Data and limits
Data defaults to ~/.hyperbench/workspace: SQLite project state plus imported document copies. Use --data-dir PATH on both the browser and MCP commands to select another location. Original repository files remain in their chosen folders. Unlinking or deleting a project never deletes those folders.
The browser binds to 127.0.0.1. Its API requires the token in the launch link and rejects foreign origins. The package has no telemetry or remote processing. Installation downloads dependencies; cloning contacts the Git remote you choose. Your agent may send selected tool results to its configured cloud model. This tool does not change the agent's privacy settings.
No code index, embeddings, automatic dependency graph, or bundled model is included. Source search reads local files on demand and reports bounded results.
PDF text extraction and page rendering are local. Scanned drawings can be inspected as images by a vision-capable agent. Automatic OCR and electrical-net reconstruction are not included.
System relationships are user-created or agent-proposed. Evidence links are references, not proof that an inferred connection is correct.
Imported document copies do not automatically follow later edits to their original files. Reimport the updated document when needed.
Repository files are read from the working tree, not immutable snapshots. Git HEAD metadata alone does not certify uncommitted content.
Only the host agent executes tests/fuzzers. No hardware flashing or physical testbench control is included.
Build from source
git clone https://github.com/metryxa-studio/hyperbench-workspace.git
cd hyperbench-workspace
npm ci
npm run build
npm startnpm run check
npm test
npx playwright install chromium
npm run test:ui
npm packnpm pack produces the same installable package used for GitHub releases, with compiled UI/server and the skills. It excludes local data, test output, and dependencies (installed by npm). See architecture and validation.
Troubleshooting
Run hyperbench doctor. If a browser tab loses its connection after a restart, use the new local link printed by hyperbench start. A missing Git credential must be fixed through your usual Git/SSH tooling. A project deleted in the UI must be removed or replaced in the corresponding agent configuration.
License
MIT. Third-party notices are in NOTICE. The Hyperbench code indexer is not part of this repository or package.
This server cannot be deployed
Maintenance
Related MCP Connectors
Project memory, semantic code search, and grounded agent context.
Agent-native notes, tasks, dev-docs, vaults, sync & handoffs. MCP + OpenAPI dual surface.
Context and control for any AI. Agent identity, scoped access, and proof across tools.
Securely search and manage workspace context files for AI agents and teams.
Related MCP Servers
- FlicenseNot gradedqualityBmaintenanceEnables AI agents to process files locally — OCR images, extract text from PDFs and DOCX, and describe images using local vision models, all without sending data to external services.-
- AlicenseNot gradedqualityBmaintenanceEnables agents to retrieve relevant context from local documents via MCP tools, supporting ingestion, semantic search, metadata filtering, and evidence inspection entirely on-device.MIT
- AlicenseAqualityAmaintenanceEnables local AI agents to perform hybrid search (BM25 + vector) over privately indexed local files, including notes, PDFs, DOCX, source code, and OCR'd images. All indexing and retrieval stay on the machine with read-only, whitelist-scoped access.3MIT
- AlicenseNot gradedqualityBmaintenanceEnables AI agents to retrieve hyper-targeted, token-efficient context from a codebase by querying a locally built AST knowledge graph, and to persist cross-session agent memory through 17 MCP tools. All indexing and retrieval run entirely offline, so proprietary code never leaves the machine.146 PyPI1MIT