CodeLens MCP

Agent-native code intelligence server with bounded workflows, precise fallback, and auditable releases.

If you are preparing automation or host configs for the eventual cutover, use the host-by-host migration guide: docs/migrate-from-codelens.md.

Pure Rust MCP server for multi-agent harnesses with hybrid retrieval (tree-sitter + semantic), mutation-gated refactoring, token compression, and enterprise-ready observability. The binary statically links SQLite, the vector store, and the ONNX runtime, so no external daemons or service installs are required for the core retrieval and mutation surfaces. Semantic search additionally needs a sidecar model directory (~80 MB ONNX) — GitHub Release tarballs bundle it automatically, but users installing via cargo install codelens-mcp must point CODELENS_MODEL_DIR at a separately-fetched model payload (see the Install Channel Matrix).

Surface Snapshot

• Workspace version: 1.13.27
• Workspace members: 3 (crates/codelens-engine, crates/codelens-mcp, crates/codelens-tui)
• Registered tool definitions: 112
• Tool output schemas: 79 / 112
• Supported language families: 30 across 49 extensions
• Profiles: planner-readonly (34), builder-minimal (39), reviewer-graph (38), evaluator-compact (14), refactor-full (54), ci-audit (46), workflow-first (22)
• Presets: minimal (27), balanced (83), full (112)
• Canonical manifest: docs/generated/surface-manifest.json

---

The Problem

Multi-agent coding harnesses fail when every agent sees too many tools, too much raw code, and too many intermediate results. Tokens get burned on tools/list, repeated file reads, and low-value raw graph expansion.

The Solution

CodeLens maintains a live, indexed understanding of your codebase and exposes it as a harness optimization layer. The model asks a precise question and gets a bounded answer with a handle for deeper expansion only when needed.

Without CodeLens                                    With CodeLens (with semantic feature on)
────────────────────────────────────────────────────────────────────────────────────────────
Read file + grep references   → 4,600 tokens       get_impact_analysis    → 1,500 tokens  (67% saved)
Read manifest + entry + files → 5,000 tokens       onboard_project        →   660 tokens  (87% saved)
Read + grep × 3 files         → 3,200 tokens       get_ranked_context     →   800 tokens  (75% saved)

Measured with tiktoken (cl100k_base) on real projects with --features semantic enabled and the bundled CodeSearchNet model loaded. Reproducible via benchmarks/token-efficiency.py. The default crates.io build (BM25 + AST only) still hits the bounded-output and workflow-shape benefits but does not run the hybrid semantic ranker.

Quick Install

Default (BM25 + AST + call-graph, no model sidecar) — works out of the box, no extra setup:

cargo install codelens-mcp

Hybrid retrieval (semantic + bundled CodeSearchNet model) — pick one:

# Option A: GitHub Release tarball — model is bundled and verified in CI.
curl -fsSL https://raw.githubusercontent.com/mupozg823/codelens-mcp-plugin/main/install.sh | bash

# Option B: cargo install with the semantic feature, then point CODELENS_MODEL_DIR at a model
#          payload (the model is not on crates.io due to the 10 MB cap).
cargo install codelens-mcp --features semantic
export CODELENS_MODEL_DIR=/path/to/codesearch/model

# Option C: Homebrew tap (macOS / Linux) — release-tarball-equivalent
brew install mupozg823/tap/codelens-mcp

HTTP daemon mode — add --features http to either path above. Source builds:

cargo install --git https://github.com/mupozg823/codelens-mcp-plugin codelens-mcp
cargo install --git https://github.com/mupozg823/codelens-mcp-plugin codelens-mcp --features semantic,http

The default cargo install codelens-mcp build was switched to default = [] in 1.10.0 (ADR-0012) so a fresh install boots without the ~80 MB ONNX sidecar. Existing users running cargo install --force will see the change in the startup banner.

Latest release: GitHub Releases. For local release comparisons, use git tag --sort=-v:refname | head -1 instead of copying a fixed tag into docs.

Install Channel Matrix

Channel	What you get	Good for	Extra install needed?
cargo install codelens-mcp	crates.io package, BM25 + AST + call-graph default (no semantic, no model needed)	Single-agent local MCP sessions, fast first install	Add --features semantic for hybrid retrieval (model sidecar required, see below). Add --features http for shared HTTP daemons.
cargo install codelens-mcp --features semantic	crates.io package + ONNX/fastembed/sqlite-vec compiled in	Hybrid retrieval users who prefer crates.io	Semantic search requires a sidecar model directory — model files are excluded from cargo publish (10 MB cap on crates.io); fetch one from a GitHub Release tarball and point CODELENS_MODEL_DIR at it
cargo install codelens-mcp --features http	crates.io package, BM25/AST default + HTTP transport	Shared daemon mode from crates.io without semantic	Combine with --features semantic,http if hybrid retrieval is wanted
GitHub Releases / installer / Homebrew	latest tagged release binary, built in CI with --features http (http,coreml on macOS)	Tagged release users who want HTTP without compiling	Model payload is bundled in the tarball and verified pre-/post-archive in CI; airgap users can rebundle via scripts/build-airgap-bundle.sh
cargo install --git ... or source build	current repository HEAD	Unreleased features on main / branch testing	Models live at crates/codelens-engine/models/codesearch/ in the source tree; no extra fetch needed

Important:

• CodeLens standalone means the codelens-mcp binary itself. Basic stdio MCP use needs only that binary plus host MCP config.
• Shared HTTP + multi-agent coordination still uses the same binary, but the binary must include the http feature and the clients must attach by URL.
• If a feature is mentioned in this repository but not present in your installed binary, compare codelens-mcp --version with the latest GitHub release and your install channel before assuming a bug.

Setup

Environment Variables

Copy .env.example to .env and fill in the values for your environment:

cp .env.example .env

Key variables include CODELENS_MODEL_DIR for semantic search, CODELENS_OTEL_ENDPOINT for telemetry, and CODELENS_PROJECT_BRIDGES_ON to opt into project-specific NL→code bridges.

Claude Code / Cursor

{
  "mcpServers": {
    "codelens": {
      "command": "codelens-mcp",
      "args": []
    }
  }
}

Shared HTTP Daemon (Multi-Agent)

Running every editor or agent as its own stdio subprocess spawns one codelens-mcp instance per session, each with its own index and embedding state. Measured on a typical developer laptop with Claude Code + Codex Desktop + Cursor attached to the same project, this adds up to 200–300 MB of duplicated resident memory for effectively the same data. The HTTP daemon collapses that into a single shared process.

If you installed from crates.io or built from source and need HTTP transport, make sure the binary was built with the http feature. The prebuilt release assets and the installer fallback should ship HTTP support.

Minimal setup:

# Start once, keep running in the background
codelens-mcp /path/to/project --transport http --profile reviewer-graph --daemon-mode read-only --port 7837

# Optional: a second daemon scoped for refactor-capable agents
codelens-mcp /path/to/project --transport http --profile refactor-full --daemon-mode mutation-enabled --port 7838

Those ports are the public generic example. In this repository's local launchd workflow, the repo-local dual-daemon installer uses :7839 for the read-only daemon and :7838 for the mutation daemon.

Every MCP client then attaches by URL instead of spawning a subprocess:

{
  "mcpServers": {
    "codelens": { "type": "http", "url": "http://127.0.0.1:7837/mcp" }
  }
}

If you are following this repository's local launchd workflow, replace the read-only example URL above with http://127.0.0.1:7839/mcp. The :7837 address remains the public generic example used throughout this section.

When to prefer HTTP vs stdio

Situation	Transport	Why
Single-agent, ephemeral sessions	stdio	Zero setup, auto-lifecycle, no port management
2+ agents (Claude + Codex + Cursor) on the same repo	HTTP	One shared index, 100–200 MB saved per extra agent
Long-running agent or automation loop	HTTP	Avoids cold-start on every session
CI / one-shot script	stdio	--oneshot matches short-lived commands
Mutation-heavy workflow needing isolation	HTTP with two daemons	Read-only port for planners, mutation-enabled port for refactor agents

For shared HTTP deployments, treat CodeLens coordination as advisory evidence rather than a central lock manager. The practical pattern is: bootstrap with prepare_harness_session, register intent with register_agent_work, claim mutation targets with claim_files, and let verify_change_readiness surface overlapping_claims as a caution signal before edits.

What the standalone binary does and does not cover:

• CodeLens only is enough for stdio use, HTTP daemon use, role-based surfaces, mutation gates, and coordination tools.
• Semantic retrieval needs the packaged model sidecar at ./models/codesearch/ next to the binary, an installed prefix sidecar such as ../models/codesearch/, or an explicit CODELENS_MODEL_DIR. Release packaging fails closed if the model payload is incomplete; release CI can point at a staged model root with CODELENS_RELEASE_MODELS_DIR. macOS release binaries enable the coreml feature so the INT8 ONNX model can use the CoreML execution provider instead of silently falling back to CPU.
• IDE adapters are external adapter endpoints that can be plugged in for IDE-specific semantic edits; no bundled adapters are required for core operation. semantic_edit_backend=tree_sitter is the default and always available.
• SCIP precise navigation needs a binary built with --features scip-backend and an external SCIP index.
• Claude -> Codex live delegation is not a CodeLens feature. It additionally needs Claude configured with a codex MCP server and a working Codex CLI install.

Recommended operating policy:

• one mutation-enabled agent per worktree
• additional agents stay planner/reviewer/read-only on the same daemon
• use codelens://activity/current to inspect active sessions, recent intent, and advisory file claims

Troubleshooting

• Failed to reconnect on the client — the daemon likely exited or the configured URL/port is wrong. Verify with curl <configured-mcp-url>; for this repository's local launchd workflow that is usually http://127.0.0.1:7839/mcp for read-only and http://127.0.0.1:7838/mcp for mutation.
• Stale index warning on first attach — expected when the watcher hasn't caught up after a daemon restart. Call refresh_symbol_index via MCP once, or restart the daemon with the project root as its CWD.
• Host config sanity check — codelens-mcp doctor <host> (or codelens-mcp status <host>) inspects the host-native files and tells you whether the CodeLens entry is attached exactly, customized, missing, or needs manual review. Add --json when another script or host automation needs a machine-readable report.
• Broken or stale ~/.local/bin/codelens-mcp — if cargo clean removed the repo build a symlink points at, or if PATH still resolves to an older cargo-installed binary that does not know newer subcommands like doctor / status, run bash scripts/sync-local-bin.sh . to rebuild and re-link the local checkout, or cargo install --path crates/codelens-mcp --force to install a fresh standalone binary under ~/.cargo/bin/.
• Multiple daemons listening on the same port — only one will actually bind; the rest exit immediately. Check the actual configured port, for example lsof -iTCP:7839 -sTCP:LISTEN or lsof -iTCP:7838 -sTCP:LISTEN in this repository's local launchd workflow.
• Health check — scripts/mcp-doctor.sh . --strict verifies that the configured transport matches an actual attach.

Auto-start on macOS (launchd)

For this repository, prefer the installer script over hand-editing plist files:

bash scripts/install-http-daemons-launchd.sh . --load

That installs two repo-local launchd agents from a current --features http,semantic build by default:

• dev.codelens.mcp-readonly -> reviewer-graph on :7839
• dev.codelens.mcp-mutation -> refactor-full on :7838

Build flag reminder (v1.10.1+): the installer builds the daemon with http,semantic by default and writes CODELENS_MODEL_DIR into the plists when the repo-local model sidecar exists. Use --no-semantic only when you intentionally want an HTTP-only daemon. See docs/release-verification.md for the full feature-flag matrix.

It also updates .codelens/config.json with repo-local host_attach URL overrides so codelens-mcp attach, status, and doctor reuse the same host-to-daemon contract.

Generic single-daemon example, if you want to hand-edit a plist instead of using the installer above:

<!-- ~/Library/LaunchAgents/dev.codelens.mcp.plist -->
<?xml version="1.0" encoding="UTF-8"?>
<plist version="1.0"><dict>
  <key>Label</key>            <string>dev.codelens.mcp</string>
  <key>ProgramArguments</key> <array>
    <string>/Users/you/.local/bin/codelens-mcp</string>
    <string>/Users/you/your-project</string>
    <string>--transport</string><string>http</string>
    <string>--profile</string><string>reviewer-graph</string>
    <string>--daemon-mode</string><string>read-only</string>
    <string>--port</string><string>7837</string>
  </array>
  <key>RunAtLoad</key>        <true/>
  <key>KeepAlive</key>        <true/>
  <key>StandardOutPath</key>  <string>/tmp/codelens-mcp.out.log</string>
  <key>StandardErrorPath</key><string>/tmp/codelens-mcp.err.log</string>
</dict></plist>

launchctl load ~/Library/LaunchAgents/dev.codelens.mcp.plist
launchctl list | grep codelens   # confirm it's running

For the separate daily aggregate audit snapshot, install the operator job with:

bash scripts/install-eval-session-audit-launchd.sh . --hour 23 --minute 55
launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/dev.codelens.eval-session-audit.codelens-mcp-plugin.plist

That scheduled job keeps JSON snapshots as the canonical history and refreshes .codelens/reports/daily/latest-summary.md plus .codelens/reports/daily/latest-gate.md after each run by default.

For an ad hoc operator snapshot without launchd, run:

bash scripts/export-eval-session-audit.sh
bash scripts/export-eval-session-audit.sh --format markdown
bash scripts/export-eval-session-audit.sh --history-summary-path .codelens/reports/daily/latest-summary.md
bash scripts/export-eval-session-audit.sh --history-gate-path .codelens/reports/daily/latest-gate.md

To summarize recent daily snapshots into a drift/trend report, run:

bash scripts/summarize-eval-session-audit-history.sh
bash scripts/summarize-eval-session-audit-history.sh --limit 7

To turn that history into an operator pass / warn / fail verdict, run:

bash scripts/eval-session-audit-operator-gate.sh
bash scripts/eval-session-audit-operator-gate.sh --fail-on-warn

If .codelens/eval-session-audit-gate.json exists, the gate script loads it automatically. CLI flags and env vars still override the repo-local policy.

The export script can also refresh that gate artifact automatically after each JSON snapshot, so scheduled operators do not need a second wrapper job just to keep latest-gate.md current.

See docs/platform-setup.md for Codex, Windsurf, VS Code, and other platforms.

Distribution Channels

Channel	Delivery	Notes
crates.io	cargo install codelens-mcp	Standard Rust install path
Homebrew tap	brew install mupozg823/tap/codelens-mcp	macOS/Linux package install
GitHub Releases	prebuilt archives	darwin-arm64, linux-x86_64, windows-x86_64
installer script	install.sh	Convenience bootstrap for release assets
source build	cargo build --release	Custom feature builds and local hacking

Why CodeLens?

	CodeLens	Read/Grep baseline
Token cost	50-87% less	Full file content every time
Context quality	Ranked, bounded, structured	Raw text, no prioritization
Multi-file impact	1 tool call	5-10 grep + read cycles
Runtime	Single Rust binary, <12ms cold start	N/A
Language support	Generated from the surface manifest	N/A
Agent awareness	Doom-loop detection, mutation gates	None

Key Features

Problem-First Workflows

Instead of starting from the full raw tool registry, begin with the workflow-first entrypoints:

Workflow	Tool	When
Explore codebase	explore_codebase	First look or targeted context search
Trace execution	trace_request_path	Follow request or symbol flow
Audit architecture	review_architecture	Boundaries, coupling, module shape
Plan safe refactor	plan_safe_refactor	Preview rename/refactor risk first
Review changes	review_changes	Diff-aware pre-merge review
Diagnose issues	diagnose_issues	File, symbol, or directory diagnosis
Cleanup duplicate logic	cleanup_duplicate_logic	Duplicate or removable logic cleanup

Role-Based Surfaces

Profile	Tools Visible	Use Case
planner-readonly	Workflow-first	Planner/architect context compression
builder-minimal	Workflow-first	Implementation with focused Codex/agent surface
reviewer-graph	Review-heavy	Graph-aware review and risk analysis
refactor-full	Preview-first + gated mutation	Safe refactors
ci-audit	Machine-oriented	CI/CD review and report emission

Adaptive Token Compression

5-stage budget-aware compression automatically adjusts response size. The per-request max_tokens parameter is honoured by the envelope budget logic; earlier versions silently capped at the profile default even when the caller explicitly asked for a larger budget.

• Stage 1 (<75% budget): Full detail pass-through
• Stage 2-3 (75-95%): Structured summarization
• Stage 4-5 (>95%): Skeleton + truncation with expansion handles

Analysis Handles

Heavy reports run as durable async jobs. Agents poll for completion and expand only needed sections:

start_analysis_job → get_analysis_job → get_analysis_section("impact")

Mutation Safety

Refactor flows require verification before code changes:

verify_change_readiness → "ready" → rename_symbol
                        → "blocked" → fix blockers first

What CodeLens Does Well (vs native grep/ls)

A complement to the existing routing matrix in CLAUDE.md. The cases below have shown up repeatedly across self-dogfood and external project sessions (see benchmarks/cl-positive-findings-2026-05-03.md for measurement notes).

Task	CodeLens	Native fallback	Why CL wins
Project bootstrap	prepare_harness_session(profile=…, detail=compact) — 1 call	ls + cargo build + python …	activation + index recovery + capability + tool surface together
Pre-merge change review	review_changes(changed_files=[…])	manual diff inspection	quantified 4-axis verifier (diagnostics / refs / tests / mutation), readiness_score 0–1, blocker_count
Function definition lookup	find_symbol(name=…)	grep -rn "def X"	exact symbol kind + signature + nearby tests in one response

For the full when not matrix (where Grep / Read win), keep the CLAUDE.md "Tool Routing — honest scenario matrix" as the single source of truth. This section only fixes the recurring "I can use CodeLens for this?" gap.

Language Support

Canonical parser families (30): C, Clojure/ClojureScript, C++, C#, CSS, Dart, Erlang, Elixir, Go, Haskell, HTML, Java, Julia, JavaScript, Kotlin, Lua, OCaml, PHP, Python, R, Ruby, Rust, Scala, Bash/Shell, Swift, TOML, TypeScript, TSX/JSX, YAML, Zig

Import-graph capable families: C, C++, C#, CSS, Dart, Go, Java, JavaScript, Kotlin, PHP, Python, Ruby, Rust, Scala, Swift, TypeScript, TSX/JSX

The canonical family/extension inventory is generated from codelens_engine::lang_registry and published in docs/generated/surface-manifest.json.

Performance

Operation	Time	Backend
find_symbol	<1ms	SQLite FTS5
get_symbols_overview	<1ms	Cached
get_ranked_context	~20ms	4-signal hybrid ranking
get_impact_analysis	~1ms	Graph cache
Cold start	~12ms	No LSP boot needed

Semantic Search

Optional embedding-based code search (feature-gated: semantic):

• Sidecar MiniLM-L12 CodeSearchNet model (ONNX INT8) — load from CODELENS_MODEL_DIR or next to the binary
• Hybrid ranking: semantic supplements structural in get_ranked_context
• 2-tier NL→code bridging: generic core (15 entries) + auto-generated project bridges (.codelens/bridges.json)
• Multi-language test symbol filtering: Python, JS/TS, Go, Java, Kotlin, Ruby

Retrieval Quality

Self-benchmark re-measured on commit 26d513e (v1.9.32, 2026-04-17), model MiniLM-L12-CodeSearchNet-INT8 (SHA256 prefix ef1d1e9c), dataset benchmarks/embedding-quality-dataset-self.json (104 queries). Two independent runs produced identical numbers (0% variance — deterministic).

Method	MRR@10	Acc@1	Acc@3	Avg ms
Lexical only (no semantic)	0.583	53%	65%	41
Semantic only	0.689	65%	74%	498
Hybrid (get_ranked_context)	0.712	68%	75%	115

Hybrid uplift over lexical: +0.128 MRR, +15% Acc@1. Semantic alone beats lexical but hybrid beats semantic by blending both signals. Identifier queries reach MRR 0.935 with every method (structural matching is sufficient); the hybrid advantage concentrates on natural-language queries (+0.159 MRR) and short phrases (+0.318 MRR).

v1.9.23 → v1.9.32 re-measurement: Hybrid −0.046 (0.758 → 0.712), Semantic −0.043, Lexical −0.018. Dataset and model unchanged. Commit span 84c825d..26d513e includes retrieval-path tuning that slightly dropped the aggregate score; the architecture refactors in v1.9.31–v1.9.32 (dispatch/, tools/, main.rs splits) do not touch retrieval code. Root-cause investigation is a follow-up in a dedicated bench session.

Cross-project matrix (6 languages, last run v1.9.23 line — not re-measured this cycle): Rust (self / axum / ripgrep), Python (django / requests), TS/JS (jest / next-js / react-core / typescript), Go (gin), Java (gson), C (curl). Historical hybrid numbers for those projects are tracked in benchmarks/embedding-quality-phase3-matrix.json.

2-tier NL→code bridges: generic core (15 entries) + auto-generated project bridges (.codelens/bridges.json).

Bridge measurement honesty (v1.9.46 three-arm ablation, 2026-04-18): on the self dataset, project bridges (.codelens/bridges.json, 659 entries) contribute 0 MRR — both-on and generic-on are bit-exact identical to six decimals. Generic core contributes +0.010 MRR overall (+0.016 on natural-language queries). Flask pilot (n=20, Python) found 0/20 generic-term matches — the generic bridges are CodeLens-dev-tooling vocabulary ("categorize", "camelcase", "who calls", "into an ast"), not a language-agnostic mapping. Cross-language bridge contribution remains unverified pending multi-repo pilots. Artifacts: benchmarks/results/v1.9.46-3arm-bridge-*.json.

Default change (v1.9.60): project bridges are now disabled by default (CODELENS_PROJECT_BRIDGES_ON=1 to re-enable) because the ablation proved zero retrieval benefit while imposing per-query file I/O. Generic bridges remain active.

# Measure on your project
python3 benchmarks/embedding-quality.py . --isolated-copy

Enterprise Features

Feature	Status
Config policy	.codelens/config.json per-project feature flags
Rate limiting	Session-level throttle (default 300 calls, configurable)
Schema versioning	schema_version: "1.0" in all responses
Intelligence sources	tree_sitter, lsp, semantic, scip — reported via get_capabilities
Mutation audit log	.codelens/audit/mutation-audit.jsonl
OTel exporter	OTLP gRPC via --features otel + CODELENS_OTEL_ENDPOINT env var
OTel-ready spans	tool.success, tool.backend, tool.elapsed_ms, otel.status_code
SBOM	CycloneDX per release
Dataset lint	CI-integrated benchmark hygiene (5 rules)
Multi-language test filter	Python, JS/TS, Go, Java, Kotlin, Ruby test symbols excluded from index
SCIP precise backend	--features scip-backend — definitions, references, diagnostics, hover
Docker	Release-runtime Dockerfile.release with healthcheck

vs Serena

Axis	CodeLens	Serena
Runtime	Single Rust binary, <12ms cold start	Python + uv
Intelligence	tree-sitter + SQLite + optional LSP/SCIP	LSP by default
Token efficiency	Bounded workflows, 50-87% savings	Standard tool responses
Workflow layer	Composite reports + analysis handles	Symbolic tools
Semantic search	Sidecar ONNX + hybrid ranking + NL bridging	No bundled model
Refactoring	Gated mutations + LSP rename/navigation/safe-delete apply	Stronger broad IDE-backed edits
Enterprise	Config policy, rate limit, OTel, SBOM	None
Offline	Works offline with a staged sidecar model	Depends on backend

See docs/serena-comparison.md for detailed gap analysis.

Building

cargo build --release                              # semantic pipeline enabled (~75MB)
cargo build --release --no-default-features        # without ML model (~58MB slim)
cargo build --release --features http              # add HTTP transport
cargo build --release --features http,coreml       # macOS HTTP + CoreML semantic runtime
cargo build --release --features otel              # add OpenTelemetry OTLP exporter
cargo build --release --features scip-backend      # add SCIP precise navigation
cargo build --release --features http,otel         # HTTP + OTel

# Core verification
cargo test -p codelens-engine
cargo test -p codelens-mcp
cargo test -p codelens-mcp --features http
cargo test -p codelens-mcp --no-default-features   # semantic=off path

Feature Flags

Feature	Description	Binary Size Impact
semantic	Semantic pipeline with sidecar ONNX model	+53MB
coreml	macOS CoreML execution provider for semantic embeddings	platform-dependent
http	Streamable HTTP + SSE transport	+2MB
otel	OpenTelemetry OTLP gRPC exporter	+4MB
scip-backend	SCIP index precise navigation	+1MB

Harness Architecture

CodeLens is designed as a harness coprocessor — it doesn't replace your agent, it makes your agent's harness smarter.

┌──────────────────────────────────────────────────────────────────┐
│                        Agent Harness                             │
│                                                                  │
│   ┌──────────┐  ┌──────────┐  ┌──────────┐  ┌──────────┐       │
│   │ Planner  │  │ Builder  │  │ Reviewer  │  │ Refactor │       │
│   └────┬─────┘  └────┬─────┘  └────┬─────┘  └────┬─────┘       │
│        │              │              │              │             │
│        └──────────────┴──────────────┴──────────────┘             │
│                              │ MCP                               │
│                    ┌─────────▼──────────┐                        │
│                    │   CodeLens MCP     │                        │
│                    │  ┌──────────────┐  │                        │
│                    │  │  Profiles    │  │ planner-readonly       │
│                    │  │  Workflows   │  │ builder-minimal        │
│                    │  │  Handles     │  │ reviewer-graph         │
│                    │  │  Gates       │  │ refactor-full          │
│                    │  └──────┬───────┘  │                        │
│                    │         │          │                        │
│                    │  ┌──────▼───────┐  │                        │
│                    │  │codelens-engine│  │ tree-sitter + SQLite  │
│                    │  │  25 langs    │  │ + embedding + graphs  │
│                    │  └──────────────┘  │                        │
│                    └────────────────────┘                        │
└──────────────────────────────────────────────────────────────────┘

Each agent role sees a different tool surface:

• Planner gets analyze_change_request, onboard_project — compressed context, no mutations
• Builder gets find_symbol, get_ranked_context — minimal surface, focused implementation
• Reviewer gets impact_report, diff_aware_references — graph-aware bounded reviews
• Refactor gets safe_rename_report, verify_change_readiness — gate-protected mutations

Harness primitives built in:

• Analysis handles — agents expand only the section they need, not the full report
• Mutation gates — verification required before code changes, preventing blind rewrites
• Doom-loop detection — identical tool calls auto-detected and redirected
• Token compression — 5-stage adaptive budget keeps responses bounded
• Suggested next tools — contextual chaining guides agents through optimal tool sequences

MCP Spec Compliance

Feature	Status
Streamable HTTP + SSE	Supported
Role-based capability negotiation	--profile flag
Tool Annotations (readOnly/destructive)	Supported
Tool Output Schemas	Generated from the surface manifest
.well-known/mcp.json Server Card	HTTP transport
HTTPS transport	Built-in rustls PEM cert/key support
Bearer/JWKS auth	Protected resource server mode
Anthropic remote connector	Tool-only compatibility profile
Analysis handles + section expansion	Supported
Durable analysis jobs	Supported
Mutation audit log	.codelens/audit/mutation-audit.jsonl
Multi-project queries	query_project
Contextual tool chaining	suggested_next_tools
MCP 2025-11-25 spec	Latest + 2025-06-18/03-26 compatibility

Quality Assurance

Suite	Gate	Scope
codelens-engine	cargo test -p codelens-engine	Parsing, ranking, embedding, IR
codelens-mcp	cargo test -p codelens-mcp	Dispatch, workflows, profiles, schemas
codelens-mcp (no semantic)	cargo test -p codelens-mcp --no-default-features	Feature-off path verification
Dataset lint	python3 benchmarks/lint-datasets.py --project .	file_exists, negative!=positive, duplicates

# Full verification
cargo test -p codelens-engine && cargo test -p codelens-mcp
cargo test -p codelens-mcp --no-default-features  # semantic=off path
python3 benchmarks/lint-datasets.py --project .     # dataset hygiene

Contributing

Contributions are welcome! Please open an issue first to discuss what you'd like to change.

# Development workflow
cargo check && cargo test -p codelens-engine && cargo test -p codelens-mcp
cargo clippy -- -W clippy::all

License

Apache-2.0