AIShell
A macOS-native MCP server that provides AI hosts with fresh workspace state, bounded context, and retained execution evidence, avoiding collapsing operations into shell strings.
README
<p align="center"> <img src=".github/og.png" alt="AIShell — Direct OS context for AI development" width="100%"> </p>
AIShell
A macOS-native MCP runtime that gives AI development hosts fresh workspace state, bounded context, and retained execution evidence—without collapsing every operation into a shell string.
AIShell owns the OS-facing state below the model: allowed roots, file identity, filesystem reconciliation, directly launched processes, complete logs, and retained artifacts. The AI host remains responsible for reasoning, threads, compaction, sub-agents, and general-purpose terminal work.
Try it in 30 seconds
Requires an Apple Silicon Mac running macOS 15 or later.
npm install -g @quolu/aishell
aishell-open
codex mcp add aishell -- /opt/homebrew/bin/aishell-mcp
In the manager app, add the folders the AI may access. Start a new Codex task and try:
Use workspace_snapshot for the initial workspace context. Run the focused tests with
run_check, and read retained output with artifact_read only if the summary omits evidence.
The default profile exposes five high-density development tools plus two always-available recovery controls:
| Tool | Purpose |
|---|---|
workspace_snapshot |
Bounded initial workspace preview, reconciled change delta, Git state, and primary context |
read_context |
Budgeted multi-file reads with SHA-256 identity and continuation |
search_context |
Budgeted search context produced by a directly launched rg worker |
run_check |
Direct process execution, primary diagnostics, and complete stdout/stderr artifacts |
artifact_read |
Range, tail, and pattern-centered reads from retained artifacts |
runtime_status |
Allowed-root, pause, worktree, and next-action state, including while paused or unconfigured |
runtime_open_manager |
Open the manager app to add roots or resume AI operations |
Why AIShell
Typical stateless integrations repeatedly ask the model to rediscover workspace state and interpret command output. AIShell keeps the stateful, OS-facing part below the model so later turns can ask for deltas and primary evidence instead of rescanning everything.
| Concern | AIShell | Typical shell-first integration |
|---|---|---|
| Workspace state | File identity plus filesystem observation and reconciliation | Re-run commands and reconstruct state from text |
| Context | Bounded, cursor-based structured results | Unbounded or manually truncated stdout |
| Execution | Executable URL, arguments, working directory, and lifecycle remain separate | A shell evaluates one command string |
| Evidence | Complete stdout/stderr retained behind expiring handles | Evidence often disappears when the response is truncated |
| Scope | Human-managed allowed roots and explicit stop state | Depends on the surrounding shell and host policy |
AIShell is not a sandbox and does not make arbitrary code execution safe. Its process rails exist to preserve typed execution and observable lifecycle—not to stop renamed binaries or child processes launched by an allowed worker.
Architecture
flowchart LR
Host[AI host<br/>reasoning · threads · compaction] --> MCP[AIShellMCP<br/>MCP 2025-11-25]
MCP --> Core[AIShellCore]
Core --> State[Allowed roots · file identity<br/>FSEvents + reconciliation]
Core --> Process[Direct process lifecycle<br/>stdout · stderr · timeout]
Core --> Evidence[Retained evidence<br/>artifacts · freshness]
Process --> Workers[git · rg · compiler · tests]
State --> macOS[macOS APIs]
Evidence --> Host
AIShellCore owns domain behavior. AIShellMCP only translates protocol requests and results. Git, ripgrep, compilers, tests, and SourceKit-LSP remain directly launched workers rather than becoming new state owners.
Install from npm
The global package adds aishell-mcp and aishell-open to PATH. aishell-open opens the bundled manager app through LaunchServices. The package runs no install script.
npm install -g @quolu/aishell
aishell-open
The current experimental build is not yet Developer ID signed or notarized.
Build from source
git clone https://github.com/kitepon-rgb/aishell.git
cd aishell
swift test
scripts/package-app.sh release
open build/AIShell.app
The MCP executable is bundled at build/AIShell.app/Contents/Helpers/aishell-mcp.
After opening the app, use Add Allowed Root to select the folders AIShell may access. A Git worktree registered under an allowed repository is recognized automatically when both sides of the worktree metadata agree.
Connect another AI host
For a Homebrew-prefix npm installation, register the absolute executable path:
codex mcp add aishell -- /opt/homebrew/bin/aishell-mcp
codex mcp get aishell
Remove the registration with:
codex mcp remove aishell
The compatibility profile retains all 25 tools. The default seven are the five development tools plus the two recovery controls; full mode adds the remaining legacy primitives:
AISHELL_TOOL_PROFILE=full /opt/homebrew/bin/aishell-mcp
The full profile includes file listing and reads, atomic SHA-256-guarded updates, copy/move/rename/Trash, direct process execution, app discovery and launch, runtime status, and manager activation.
Execution and safety boundaries
- AIShell never evaluates a shell command string. It resolves a development program from
PATHto an executable URL and keeps arguments, environment, and working directory separate. - Direct launch of shell and wrapper basenames such as
sh,bash,zsh,env, andosascriptis rejected as a product rail, not advertised as a security boundary. run_checkis an open-world capability: an allowed worker may update files, launch child processes, or access the network. AI hosts may require approval before execution.- Text updates may use SHA-256 or expected old text as a precondition. Deletes go to Trash.
- The manager app can stop normal operations globally. Runtime status and manager activation remain available while stopped.
Current limitations
- The stdio server handles one request at a time.
- MCP cancellation and concurrent run polling are not implemented.
- A timeout terminates the directly owned process tree, but an allowed worker remains capable of open-world side effects before termination.
- Initial workspace entries are a bounded preview; later deltas are cursor-paged.
- Developer ID signing and notarization are not yet configured.
Development
swift test
scripts/package-app.sh release
Run xcodegen generate to regenerate AIShell.xcodeproj. The authoritative implementation lives under Sources/AIShellCore, Sources/AIShellMCP, and Sources/AIShellApp; focused tests live under Tests/.
<details> <summary>Local Xcode verification note</summary>
On the original verification machine, Xcode 26.6 and the installed CoreSimulator build version did not match, so xcodebuild stalled before XCBuild began. The same Swift 6.3.3 toolchain passed through SwiftPM. That host issue was not counted as source success.
</details>
Contributing and security
See CONTRIBUTING.md before proposing a change. Please report vulnerabilities through the private process in SECURITY.md, not through a public issue.
Release notes are kept in docs/. GitHub Releases are the public record for shipped versions.
License
AIShell is licensed under the Apache License 2.0.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.