codescout MCP server
Commit Nov 10, 2026Gives AI coding agents symbol navigation, semantic search, persistent memory, and artifact indexing across multi-project workspaces.
Also known as mareurs/codescout
- Stars
- 27
- Tools
- —
- Downloads, 30 days
- —
- Latest version
- —
- Last commit
- Nov 10, 2026
- License
- MIT
What is codescout MCP?
codescout gives AI coding agents symbol navigation, semantic search, persistent memory, and artifact indexing across multi-project workspaces.
Summary generated from the project's README
About codescout MCP server
Providing IDE-grade intelligence to AI coding agents is what codescout enables through advanced symbol navigation and semantic search. It functions as a local MCP server that connects to agents like Claude Code, GitHub Copilot, and Cursor, allowing them to navigate complex codebases without reading entire files. By using LSP-backed tools such as symbols, references, and call graphs, it provides precise context across nine supported languages including Rust, Python, and TypeScript.
Developers can manage multi-project workspaces using a workspace.toml configuration to maintain per-project memory and indexing. The server includes a librarian component that indexes markdown artifacts like specs, plans, and ADRs into a SQLite catalog. For deep semantic retrieval, it supports a retrieval stack using a local ONNX model or an external Qdrant setup with a reranker. It is designed for token efficiency, providing compact details on demand rather than dumping large amounts of text into the agent's context window.
What you can do with it
- 1Navigate complex code structuresAn agent can find specific function definitions or call sites using symbol navigation instead of searching through raw text.
- 2Retrieve project documentationA developer can ask an AI to find all active specifications related to a specific part of the retrieval pipeline.
- 3Perform concept-based code searchesUsers can locate code by its underlying meaning using semantic search powered by local or external models.
- 4Manage multi-project contextDevelopers can register related projects to allow an AI to navigate across different repositories within a single workspace.
Written from the project's README. Every use case is checked against a quote from it.
Automated checks
Ran Nov 14, 2026Full report →Not verified yet. A server needs a score of 80, no serious finding, and enough data for us to check. How checks work · What to fix for more points
- Source repository linked and reachablemareurs/codescout is public and reachablegithub.com/mareurs/codescout ↗Pass
- Package names this repository as its sourceNot checked
- Build provenance (Sigstore attestation)Not checked
- Open source license (OSI)MITspdx.org/licenses/MIT.html ↗Pass
- Commit or release in the last yearLast commit or release 2026-11-10github.com/mareurs/codescout ↗Pass
- OpenSSF ScorecardNot checked
- Package name not close to a popular packageNot checked
- No critical advisory in the package or its dependenciesNot checked
- No hidden instructions to the model0 tool descriptions and the README read; no override phrases, credential paths or hidden charactersPass
- Destructive tools are labelledNot checked
- Remote endpoint answers the MCP handshakeNot checked
README
View on GitHubcodescout
MCP server giving AI coding agents IDE-grade code intelligence — symbol navigation, semantic search, persistent memory, workspace-wide artifact tracking — optimized for token efficiency.
📖 Full manual at mareurs.github.io/codescout — installation, agent integrations, every tool, every concept.
Works with Claude Code, GitHub Copilot, Cursor, and any MCP-capable agent.
What it does
- Symbol navigation —
symbols,references,symbol_at,call_graph,edit_code, backed by LSP across 9 languages - Semantic search — find code by concept using a local ONNX model or an external Qdrant + reranker stack
- Library navigation — explore dependency source with scoped search and version tracking
- Artifact tracking (new) — index, query, and link the markdown around your code: specs, plans, ADRs, runbooks, memories. See Artifacts below.
- Multi-project workspaces — register related projects in
workspace.tomlfor cross-project navigation with per-project memory and indexing - Token efficiency — compact by default, details on demand, never dumps full files
Why not just read files?
| Without codescout | With codescout |
|---|---|
| Agent reads full files to find one function | Navigates by symbol name — zero file reads |
grep returns noise (comments, strings, docs) |
references returns exact call sites |
| Context burns on navigation overhead | Token-efficient by design |
| State lost between sessions | Persistent memory + artifact catalog across sessions |
| Re-reads same modules from different entry points | Symbol index built once, queried instantly |
Quick Start
cargo build
./target/debug/codescout start --project /path/to/code
Register as an MCP server in ~/.claude/settings.json:
{
"mcpServers": {
"codescout": {
"command": "codescout",
"args": ["start", "--project", "."]
}
}
}
Onboarding is essential. Run
onboarding()once per project on first connect — it discovers languages, reads key files, and generates a project-specific system prompt and memory files. Without it, the agent navigates blind.Tip: Install the codescout-companion plugin to auto-steer Claude toward codescout tools in every session — including subagents.
→ Full setup, agent configs, and project-toml options: mareurs.github.io/codescout
Artifacts
codescout now embeds librarian — a workspace-wide artifact registry that indexes the markdown around your code (specs, plans, ADRs, runbooks, memories, audits, handoffs, roadmaps, user docs), stores metadata + a typed link graph in SQLite, and exposes it to the agent as queryable structured data instead of opaque file listings.
Why it matters. The code is half the project. The other half lives in
markdown files scattered across docs/, design notes, and runbooks. Without
indexing them, the agent re-discovers project history every session. With
librarian, "find every active spec touching the embedder" is a single tool call.
How it works.
- Every markdown file is classified by frontmatter or path-rule into a kind
(spec / plan / adr / runbook / memory / tracker / doc / ...) and indexed
into a SQLite catalog at
~/.local/share/librarian/catalog.db. - A typed link graph (
supersedes,implements,references, ...) connects artifacts; walk it transitively withartifact(action="graph"). - A 4-tier scope ladder (
project→repo→umbrella→all) lets the agent widen beyond the current sub-project on demand. - Optional semantic search via
LIBRARIAN_EMBED_*turnsartifact(action="find", semantic="…")into vector search.
How to use it.
# Find every active spec touching the retrieval stack
artifact(action="find", kind="spec", status="active", semantic="retrieval pipeline")
# Read one, with its outgoing link graph
artifact(action="get", id="abc123", include_links=true)
# Connect a plan to the spec it implements
artifact(action="link", src_id="<plan>", dst_id="<spec>", rel="implements")
# Pack a topic neighbourhood into a single markdown bundle for context
librarian(action="context", topic="hybrid sparse + dense retrieval", max_tokens=4000)
Trackers — augment any artifact with a persistent prompt + params, and codescout auto-refreshes it on demand. Friction logs, observation tables, sprint roadmaps — written once, maintained across sessions.
Embedded by default. No separate server to install. Opt out with
LIBRARIAN_ENABLED=0 for a leaner tool surface.
→ Full artifact guide: librarian embedded · trackers · time-travel · augmentation templates
Retrieval Stack
semantic_search runs through an external Docker Compose stack (Qdrant +
dense embedder + SPLADE sparse + cross-encoder rerank). Two profiles: cpu
and gpu.
docker compose --profile cpu --env-file .env.cpu up -d
# or --profile gpu --env-file .env.gpu
set -a; source .env.cpu; set +a
cargo run --release --bin sync_project -- . codescout
| Profile | Initial sync (~18k chunks) | Sustained throughput |
|---|---|---|
| cpu (llama-server, 4 threads) | ~125 min | 2.4 chunks/s |
| gpu (llama-server-cuda, RTX A5000) | ~2.6 min | 117-132 chunks/s |
→ Stack setup, model choices, and tuning
Agent integrations
| Agent | Guide |
|---|---|
| Claude Code | mareurs.github.io/codescout/agents/claude-code |
| GitHub Copilot | mareurs.github.io/codescout/agents/copilot |
| Cursor | mareurs.github.io/codescout/agents/cursor |
Multi-agent infrastructure
codescout's design is informed by research on compound error in multi-agent systems — failure rates of 41–87% in production pipelines. This drove the choice of single-session skill-based workflows over agent orchestration chains. Read the analysis →
Guidance that agents actually follow
codescout's prompt surfaces — Iron Laws, get_guide topics, tracker rendering,
result envelopes — are shaped by measured findings on how coding agents trust,
follow, and resist the content a tool surfaces. The short version: authority framing
(a persona, a "trusted" marker) buys nothing; placement and server-computed
provenance buy everything. Trust attaches to what the server computes
(symbols, git state, freshness keys), never to a marker the content carries about
itself — because anything in a file body is forgeable by whoever wrote the file.
Read the findings →
Kotlin & Rust LSP multiplexers
Kotlin and Rust LSPs are expensive to boot and allow only one process per
workspace. codescout ships per-language multiplexers — a detached codescout mux process shares one LSP across all codescout instances. Cold-start happens
once; subsequent sessions connect instantly. No configuration needed.
| Metric | Without mux | With mux |
|---|---|---|
| LSP processes per machine | 1 per session (~2GB each) | 1 shared |
| Cold start on 2nd session | 8–15s | ~0s |
→ Kotlin LSP multiplexer · Rust
Tools
Symbol navigation · File operations · Shell · Semantic search · Memory · Library navigation · Artifacts · Workflow & Config
Supported languages: Rust, Python, TypeScript/JavaScript, Go, Java, Kotlin, C/C++, C#, Ruby.
Experimental Features
New features land on experiments before reaching master. They may change
or be removed without notice.
→ Browse experimental features
Contributing
See CONTRIBUTING.md. PRs from Claude Code are welcome.
License
codescout MCP questions
What languages does codescout support for symbol navigation?
It supports Rust, Python, TypeScript/JavaScript, Go, Java, Kotlin, C/C++, C#, and Ruby through LSP-backed symbol tools.
Do I need to run a separate server for code search?
Semantic search can use a local ONNX model, or an external Docker Compose stack with Qdrant, a sparse embedder, and a reranker. The artifact registry is embedded and needs no separate server.
What does the artifact tracking feature do?
It indexes markdown files like specs, plans, ADRs, and runbooks into a SQLite catalog with a typed link graph, so the agent can find and connect them in single tool calls.
How do I set up codescout with Claude Code?
Register it as an MCP server in your settings file, then run onboarding() once per project so the agent can discover languages and key files.
Does codescout work with a multi-project workspace?
Yes; you register related projects in workspace.toml to get cross-project navigation with per-project memory and indexing.