XuanLing MCP

English | Simplified Chinese

XuanLing MCP is a cross-platform, local Model Context Protocol server for coding agents. It exposes 42 typed tools over stdio for filesystem work, process execution, project discovery, durable memory, artifacts, and long-running sessions.

The server is designed for hosts that need deterministic schemas, structured failures, explicit filesystem capabilities, and memory writes that cannot silently become canonical state.

Highlights

• Typed filesystem operations with strict edits, SHA-256 preconditions, paginated search, resumable reads, and explicit output budgets.
• Proposal-first Memory v2 with immutable record versions, explicit review, project/workspace scope isolation, and deterministic lexical recall.
• Direct process execution using program + argv without an implicit shell, with cancellation applied to the descendant process tree.
• Selective tool profiles so each host can expose only the capability families it needs.
• Native npm distribution for macOS, Linux, and Windows with no postinstall compilation or remote binary download.
• Stable MCP contracts backed by protocol, golden, persistence, restart, and cross-platform test suites.

Installation

The npm launcher requires Node.js 18.17 or newer and installs the matching native binary for the current platform.

npm install --global @xuanling-rs/xuanling-mcp@0.4.0
xuanling-mcp --version

An MCP client can pin the same version without a global installation:

{
  "mcpServers": {
    "xuanling": {
      "command": "npx",
      "args": [
        "-y",
        "@xuanling-rs/xuanling-mcp@0.4.0",
        "--workspace-root",
        "/absolute/path/to/project",
        "--tool-profile",
        "core",
        "--tool-profile",
        "fs",
        "--tool-profile",
        "memory"
      ]
    }
  }
}

Pinning the package version keeps the discovered MCP schema stable for an active project. A global or project-local installation avoids the package resolution step when the host starts the server frequently.

Supported platforms

Unsupported OS, CPU, and libc combinations fail with an explicit launcher error. The package does not compile Rust or download executables during installation. The launcher validates package metadata and the native binary's SHA-256 before starting the server.

Build from source

Source builds require Rust 1.98.

cargo build --locked --release -p xuanling-mcp
./target/release/xuanling-mcp --workspace-root /absolute/path/to/project

Tool Profiles

The default catalog contains all 42 tools. Repeat --tool-profile to combine smaller, stable groups:

all takes precedence when combined with another profile. Discovery and dispatch use the same selection, so a hidden tool cannot be called by name.

Catalog transport

tools/list returns the selected static catalog in pages of eight tools. Follow the opaque nextCursor until it is absent; a malformed, stale, or cross-catalog cursor fails with JSON-RPC -32602 and data.reason: "invalid_cursor" instead of restarting from page one. Initialization metadata publishes xuanling.tool_count for the complete filtered catalog and xuanling.catalog_sha256 as a stable digest of every model-visible definition.

The catalog does not mutate during a server process, so XuanLing deliberately does not advertise tools.listChanged. Pagination bounds MCP transport frames; whether the complete catalog enters model context remains a Host projection decision.

Filesystem Safety

Filesystem access is unrestricted when neither capability flag is supplied. Production host configurations should declare at least one root:

• --workspace-root <PATH> is repeatable and grants read/write/delete access plus child-process working-directory admission inside the root.
• --read-root <PATH> is repeatable and grants read/list/search/hash access while rejecting writes, deletion, and child-process working directories.
• Supplying only --read-root creates a read-only deployment.

Mutating tools support explicit preimage checks. Use expected_sha256 for whole-file replacement and exact edits, or expected_preimage_sha256 for fs_patch. A conflict is reported before the write when the file changed after it was read.

Window-capable tools accept an explicit output selector such as {"mode":"bounded","max_bytes":65536}. Truncated reads and searches return a typed cursor or resume token; they do not silently discard the remaining result.

The filesystem capability controls paths opened by XuanLing. It is not an OS sandbox for arbitrary child programs: host approval and process isolation remain responsibilities of the MCP host and its execution environment.

Memory v2

Memory v2 separates proposals from canonical records:

1. memory_search and memory_get inspect active records.
2. memory_candidate_create, memory_candidate_replace, or memory_candidate_archive creates a pending proposal.
3. memory_review accepts or rejects a specific proposal revision. Only an accepted review atomically advances the canonical record head.
4. Immutable versions, terminal reviews, and append-only feedback retain the history needed for audit and deterministic recovery.

Scopes use strict tagged values for global, project, and workspace. Ancestor search follows workspace -> project -> global only when requested; sibling projects are never searched.

Recall uses a deterministic lexical query plan over SQLite FTS5 (unicode61 and trigram), multi-channel fusion, visibility filtering, and stable reranking. The default release does not require or download an embedding model and does not perform network access for recall.

The default database is ~/.xuanling/memory.db. Override it with --memory-db <PATH> and optionally provide --default-namespace <VALUE>. The stdio server resolves and opens that database only when the first valid Memory tool is called; initialize, discovery, non-memory tools, and malformed Memory requests do not create or migrate it. Memory initialization failure does not disable non-memory tools; memory calls return a structured unavailable error instead.

Maintenance

Canonical data can be exported, imported into an empty database, and used to rebuild the derived search projection:

xuanling-mcp --memory-db /path/to/memory.db memory export --output backup.jsonl
xuanling-mcp --memory-db /path/to/empty.db memory import --input backup.jsonl
xuanling-mcp --memory-db /path/to/memory.db memory rebuild-index

Export writes a versioned JSONL stream with counts and a SHA-256 trailer. Import validates the complete stream before one transactional write, and rebuild-index never changes canonical rows.

DeepSeek Harness

Conversational install from the repository URL

Copy https://github.com/umbrella22/xuanling into a DeepSeek Harness (DSH) chat and ask it to install the XuanLing DSH integration. DSH will read the repository-owned installer Skill, ask which profile and preset to use, show the exact frozen npm version and package changes for confirmation, then install and verify them through dsh plugin.

This is a model-orchestrated workflow: DSH itself does not install arbitrary URLs. When needed, the agent obtains the fixed repository ref in a new temporary checkout, uses it only for bounded document discovery, and removes the checkout before asking the first question. Path discovery may list tracked paths or run a locator whose model-visible output is path names only; source and manifest bodies never enter model context. The only loaded content is an allowlisted root README, the installer Skill, and optionally the DSH integration guide. The agent never executes repository code or installs from the checkout; profile packages still come only from the public npm registry. The manual integration guide below remains the fallback when interactive questions or repository access are unavailable.

integrations/deepseek-harness contains host-specific bundles for additive Memory tools, additive or replacement filesystem tools, schema projection, strict overwrite policy, and two on-demand workflow Skills. The integration remains outside the Rust tool contracts, allowing DeepSeek Harness-specific routing and policy to evolve without changing the MCP catalog for other hosts.

The shipped DSH runtime bundles use Host-side lazy projection: they cache every MCP catalog page but initially expose only mcp_catalog__xuanling. The model searches that compact control and activates exact raw names before ordinary mcp__xuanling__* calls. This reduces initial schema cost without treating MCP pagination or list_changed as capability selection.

See the DeepSeek Harness integration guide for bundle selection, installation, and runtime configuration.

Repository Layout

The current documentation index is available at docs/README.md. Repository provenance and the detached workspace boundary are recorded in docs/repository-boundary.md.

Development

Repository development requires Rust 1.98, Node.js 22.14 or newer, and npm 11.5.1 or newer.

cargo fmt -p xuanling-toolkit -p xuanling-memory -p xuanling-mcp -- --check
cargo check -p xuanling-toolkit -p xuanling-memory -p xuanling-mcp --all-targets
cargo clippy -p xuanling-toolkit -p xuanling-memory -p xuanling-mcp --all-targets -- -D warnings

cargo test -p xuanling-toolkit --features test-fixtures --test contract
cargo test -p xuanling-memory --test contract
cargo test -p xuanling-mcp --test protocol
cargo test -p xuanling-mcp --test golden

npm --prefix npm run check
npm --prefix npm run check:docs
npm --prefix npm test

# Non-gating, single-process benchmark for the complete paginated catalog.
cargo build --locked --release -p xuanling-mcp
npm --prefix npm run benchmark:catalog -- --binary target/release/xuanling-mcp

The complete host contract and error mapping are documented in the MCP integration guide. npm package assembly and publishing are documented in npm/README.md.

License

Licensed under the MIT License.