codebase-memory-mcp

The fastest and most efficient code intelligence engine for AI coding agents. Full-indexes an average repository in milliseconds, the Linux kernel (28M LOC, 75K files) in 3 minutes. Answers structural queries in under 1ms. Ships as a native executable with a small verified runtime-asset set for macOS, Linux, and Windows — download, run install, done.

High-quality parsing through tree-sitter AST analysis across all 162 languages, enhanced with Hybrid LSP semantic type resolution for Python, TypeScript / JavaScript / JSX / TSX, PHP, C#, Go, C, C++, Java, Kotlin, Rust, and Perl — producing a persistent knowledge graph of functions, classes, call chains, HTTP routes, and cross-service links. 15 MCP tools. No language runtime, hosted service, or API key. Plug and play across 45 supported automatic/conditional client surfaces.

Research — The design and benchmarks behind this project are described in the preprint Codebase-Memory: Tree-Sitter-Based Knowledge Graphs for LLM Code Exploration via MCP (arXiv:2603.27277). Evaluated across 31 real-world repositories: 83% answer quality, 10× fewer tokens, 2.1× fewer tool calls vs. file-by-file exploration.

Security & Trust — This tool reads your codebase and writes to your agent configuration files. That is what it is designed to do. If you prefer to audit before running, the full source is here. For each release product, three behaviourally identical executable candidates (unstripped, debug-stripped, stripped) are submitted to VirusTotal before testing; the selected candidate is then packaged with its SHA-256 unchanged. Release notes link every measured candidate result. Publication permits only the narrowly documented single-Microsoft !ml tolerance in SECURITY.md. All processing happens 100% locally; your code never leaves your machine. Found a security issue? We want to know — see SECURITY.md. Security is Priority #1 for us.

Why codebase-memory-mcp

• Extreme indexing speed — Linux kernel (28M LOC, 75K files) in 3 minutes. RAM-first pipeline: LZ4 compression, in-memory SQLite, fused Aho-Corasick pattern matching. Memory released after indexing.
• Plug and play — native executable plus authenticated release-owned assets for macOS (arm64/amd64), Linux (arm64/amd64), and Windows (amd64). The native install needs no Docker, language runtime, or API keys. Download → install → restart agent → done.
• 162 languages — vendored tree-sitter grammars compiled into the binary. Nothing to install, nothing that breaks.
• 120x fewer tokens — 5 structural queries: ~3,400 tokens vs ~412,000 via file-by-file search. One graph query replaces dozens of grep/read cycles.
• 45 supported automatic/conditional client surfaces — install configures detected clients and safely activates conditional clients only when their documented platform, marker, or explicit existing config path is present. See Multi-Agent Support for the complete matrix and manual/UI-only boundaries.
• Built-in graph visualization — 3D interactive UI at localhost:9749, served from the binary itself.
• Infrastructure-as-code indexing — Dockerfiles, Kubernetes manifests, and Kustomize overlays indexed as graph nodes with cross-references. Resource nodes for K8s kinds, Module nodes for Kustomize overlays with IMPORTS edges to referenced resources.
• 15 MCP tools — search, trace, architecture, impact analysis, targeted index-coverage checks, Cypher queries, dead code detection, cross-service HTTP linking, ADR management, and more.

Quick Start

One-line install (macOS / Linux):

curl -fsSL https://raw.githubusercontent.com/DeusData/codebase-memory-mcp/main/install.sh | bash

With graph visualization UI:

curl -fsSL https://raw.githubusercontent.com/DeusData/codebase-memory-mcp/main/install.sh | bash

Windows (PowerShell):

# 1. Download the installer
Invoke-WebRequest -Uri https://raw.githubusercontent.com/DeusData/codebase-memory-mcp/main/install.ps1 -OutFile install.ps1

# 2. (Optional but recommended) Inspect the script
notepad install.ps1

# 3. Unblock the downloaded file (removes Mark-of-the-Web restriction added by browsers/Invoke-WebRequest)
Unblock-File .\install.ps1

# 4. Run it
.\install.ps1

Note: If you see a script execution policy error, run Set-ExecutionPolicy -Scope Process Bypass first, or invoke with PowerShell -ExecutionPolicy Bypass -File .\install.ps1.

Options: --skip-config (binary only, no agent setup), --dir=<path> (custom location).

Antivirus note: Microsoft Defender may flag a release binary as Trojan:Script/Wacatac.B!ml. This is a known false positive — typically 61 of ~62 engines return clean, and the same detection family hits gh, llama.cpp, Godot and Microsoft's own Go toolchain. See Antivirus False Positives for the evidence, how to verify the artifacts yourself, and how to report it if you think we are wrong.

Restart your coding agent. Say "Index this project" — done.

1. Download the archive for your platform from the latest release:

   • codebase-memory-mcp-<os>-<arch>.tar.gz (macOS/Linux) or .zip (Windows)

2. Extract and install (each archive includes install.sh or install.ps1):

   macOS / Linux:

   tar xzf codebase-memory-mcp-*.tar.gz
   ./install.sh
   

   Windows (PowerShell):

   Expand-Archive codebase-memory-mcp-windows-amd64.zip -DestinationPath .
   Unblock-File .\install.ps1
   .\install.ps1
   

3. Restart your coding agent.

The install command automatically strips macOS quarantine attributes and ad-hoc signs the binary — no manual xattr/codesign needed.

The install command auto-detects installed coding agents and configures their documented MCP entries plus durable instructions, skills, and lifecycle hooks where supported.

Session Coordination Daemon

CBM automatically shares one per-account coordination daemon across Claude Code, Codex, OpenCode, and every other configured client. There is no opt-in setting for MCP servers or hook clients: the first daemon-backed CBM session starts it, each session registers its own work, and the final session shuts it down. The daemon owns long-lived background services such as watchers, shared indexing jobs, and the optional UI. Closing one session cancels work owned only by that session, while work still needed by another session continues.

The detached daemon does not depend on an MCP frontend's stderr. It keeps owner-only durable records under the canonical ${CBM_CACHE_DIR}/logs directory (default ~/.cache/codebase-memory-mcp/logs):

Thin frontends still write immediate startup and session-specific errors to their own stderr; MCP JSON-RPC stdout remains clean.

All active CBM processes must run the exact same version, executable build, coordination ABI, and canonical cache root. Equivalent CBM_CACHE_DIR aliases resolve to the same root; a genuinely different root is rejected while any CBM process is active. MCP servers, hooks, one-shot CLI commands, temporary index workers, and the daemon share a crash-safe OS admission barrier; starting an ordinary conflicting process fails before doing work and records an explicit conflict in ${CBM_CACHE_DIR}/logs/daemon-conflicts.ndjson.

The native install, update, and uninstall commands are the deliberate exception to that conflict rule. Download, verification, and private same-filesystem staging happen first so a bad candidate never disrupts active work. Activation then publishes account-wide maintenance intent, asks the daemon and every temporary local operation to cancel, and waits to a finite deadline for all coordinated CBM processes to exit. It holds the admission and lifetime barriers exclusively while changing the active binary, configuration, PATH, or indexes. New CBM work cannot enter during this window. Activation progress and results are recorded in ${CBM_CACHE_DIR}/logs/activation-events.ndjson, and a successful command tells you to restart open coding-agent sessions so they launch the activated build.

Package-manager setup (npm, PyPI, or Go) verifies and publishes a coherent private cached runtime set. Sidecars are replaced before the executable with per-file atomic renames; an interrupted multi-file publication is detected and repaired on the next launch rather than being described as one crash-atomic filesystem transaction. It does not replace the active native installation and therefore does not stop running CBM sessions. When that cached binary is executed, it still enters the same exact-build admission barrier. The shell and PowerShell installers invoke the verified candidate's native install command, so they do receive the full account-wide activation guarantee.

The ordinary cli mode is intentionally separate: it runs one command locally and never starts or connects to the coordination daemon, registers a daemon session, or starts watchers/UI. Its only shared state is the OS admission barrier plus per-project locks for graph mutations. While the command is running, a temporary monitor lets activation cancel that operation and its supervised worker safely; the monitor exits with the command and never becomes a standing daemon. See CLI Mode for details.

Graph Visualization UI

The graph UI is built into the binary — every install on every channel has it. Then run it:

codebase-memory-mcp --ui=true --port=9749

Open http://localhost:9749 in your browser. The UI is owned by the shared coordination daemon, so concurrent agent sessions do not start duplicate HTTP servers.

Auto-Index

Enable automatic indexing on MCP session start:

codebase-memory-mcp config set auto_index true

When enabled, new projects are indexed automatically on first connection. Previously-indexed projects are registered with the background watcher for ongoing git-based change detection. Configurable file limit: config set auto_index_limit 50000.

Watcher registration is controlled separately by auto_watch (default true). Set config set auto_watch false to keep a session from registering its project with the background watcher — useful when working across many projects and you want each session contained to explicit indexing.

To turn the watcher off entirely, set config set watcher_enabled false (default true): the background poll thread never starts and no project is registered, while auto_index and manual index_repository keep working. Unlike auto_watch — which is consulted per session — watcher_enabled is read once when the background daemon starts, so run codebase-memory-mcp daemon stop after changing it; reconnecting your MCP client alone will not restart the daemon. See docs/CONFIGURATION.md.

Keeping Up to Date

Updates run from the install script on every platform, not from inside the running binary. codebase-memory-mcp update validates your flags and then prints the exact command to run:

# macOS / Linux
bash "<install-dir>/install.sh"

# Windows
powershell -ExecutionPolicy Bypass -File "<install-dir>\install.ps1"

The install script is placed next to the binary at install time, so the printed path resolves beside the executable. It is idempotent, so re-running it is the update: it stops the daemon, retires the running binary, installs the new one, and cleans up.

Why it works this way. On Windows it is a hard requirement — a running executable cannot replace its own image, so the swap has to happen from a process that is not the binary being replaced. On macOS and Linux it is a deliberate choice: an in-process updater is structurally a downloader (fetch an archive, verify it, unpack it, mark a file executable, run it), and shipping that composite in every binary to serve a command most people run a handful of times is a poor trade. The release archives now carry no download URLs at all, and cbm makes no network request of its own accord — it does not check for new versions in the background, and nothing phones home. You find out about releases from the install script, your package manager, or GitHub.

If PowerShell refuses to run the script because the file came from the internet, Unblock-File it first.

Installed through npm or pip? Update with your package manager on every platform (npm install -g codebase-memory-mcp@latest / pip install -U codebase-memory-mcp).

Uninstall

codebase-memory-mcp uninstall

Removes owned agent config entries, skills, hooks, instructions, and the installed binary. Existing graph indexes are listed and deleted only after confirmation.

The install script placed beside the binary is reported, not deleted — uninstall prints its path and the rm command for it. It is left alone on purpose: it may be your own copy, a symlink into a checkout, or managed by a package manager, and an uninstaller should not delete a file it cannot prove it owns.

Features

Graph & analysis

• Architecture overview: get_architecture returns languages, packages, entry points, routes, hotspots, boundaries, layers, and clusters in a single call
• Architecture Decision Records: manage_adr persists architectural decisions across sessions
• Louvain community detection: Discovers functional modules by clustering call edges
• Git diff impact mapping: detect_changes maps uncommitted changes to affected symbols with risk classification
• Call graph: Resolves function calls across files and packages (import-aware, type-inferred)
• Dead code detection: Finds functions with zero callers, excluding entry points
• Cypher-like queries: MATCH (f:Function)-[:CALLS]->(g) WHERE f.name = 'main' RETURN g.name

Search

• Semantic search (semantic_query): vector search across the entire graph, powered by bundled Nomic nomic-embed-code embeddings (40K tokens, 768d int8) compiled into the binary — no API key, no Ollama, no Docker. 11-signal combined scoring (TF-IDF, RRI, API/Type/Decorator signatures, AST profiles, data flow, Halstead-lite, MinHash, module proximity, graph diffusion).
• BM25 full-text search via SQLite FTS5 with cbm_camel_split tokenizer (camelCase / snake_case aware)
• Structural search (search_graph): regex name patterns, label filters, min/max degree, file scoping
• Code search (search_code): graph-augmented grep over indexed files only

Cross-service linking

• HTTP route ↔ call-site matching with confidence scoring
• gRPC, GraphQL, tRPC service detection with protobuf Route extraction
• Channel detection (EMITS / LISTENS_ON) for Socket.IO, EventEmitter, and generic pub-sub patterns across 8 languages with constant resolution

Cross-repo intelligence

• CROSS_* edges link nodes across multiple repos indexed under the same store
• Multi-galaxy 3D UI layout for cross-repo architecture visualization
• Cross-repo architecture summary combining services, routes, and dependencies across the indexed fleet

Edge types (selected)

• CALLS — a callable is invoked at the source site
• CALL_REFERENCE — a callable is used at a supported reference site (for example, a direct value argument) and resolves to one exact target
• USAGE — an identifier is used, but a unique callable target is not proven (including ambiguous or complex expressions)
• IMPORTS, DEFINES, IMPLEMENTS, INHERITS
• HTTP_CALLS, ASYNC_CALLS (cross-service)
• EMITS, LISTENS_ON (channels)
• DATA_FLOWS with arg-to-param mapping + field access chains
• SIMILAR_TO (MinHash + LSH near-clone detection, Jaccard scored)
• SEMANTICALLY_RELATED (vocabulary-mismatch, same-language, score ≥ 0.80)

Indexing pipeline

• 158 vendored tree-sitter grammars compiled into the binary
• Generic package / module resolution — bare specifiers like @myorg/pkg, github.com/foo/bar, use my_crate::foo resolved via manifest scanning (package.json, go.mod, Cargo.toml, pyproject.toml, composer.json, pubspec.yaml, pom.xml, build.gradle, mix.exs, *.gemspec)
• Infrastructure-as-code indexing — Dockerfiles, Kubernetes manifests, Kustomize overlays as graph nodes
• Hybrid LSP semantic type resolution for Python, TypeScript / JavaScript / JSX / TSX, PHP, C#, Go, C, C++, Java, Kotlin, Rust, and Perl — a lightweight C implementation of language type-resolution algorithms, structurally inspired by and compatible with major language servers including tsserver / typescript-go, pyright, gopls, Roslyn, Eclipse JDT, and rust-analyzer (parameter binding, return-type inference, generic substitution, JSX component dispatch, JSDoc inference for plain JS files, namespace + trait + late-static-binding resolution for PHP, file-scoped namespaces + records + LINQ method syntax for C#, class-hierarchy + overload + lambda resolution for Java, extension-function + scope-function resolution for Kotlin, trait-method + UFCS resolution for Rust)
• RAM-first pipeline: LZ4 compression, in-memory SQLite, single dump at end. Memory released after.

Distribution & operation

• Native runtime set, zero infrastructure services: SQLite-backed, persists to ~/.cache/codebase-memory-mcp/
• Auto-sync: Background watcher detects file changes and re-indexes automatically
• Route nodes: REST endpoints are first-class graph entities
• CLI mode: codebase-memory-mcp cli search_graph '{"project": "my-project", "name_pattern": ".*Handler.*"}'
• Available on: npm, PyPI, Homebrew, Scoop, Winget, Chocolatey, AUR, go install

Team-Shared Graph Artifact

Commit a single compressed file to your repo and your teammates skip the reindex.

.codebase-memory/graph.db.zst is a zstd-compressed snapshot of the knowledge graph that lives next to your source. When you index, the artifact is written or refreshed; when a teammate clones the repo and runs codebase-memory-mcp for the first time, the artifact is decompressed and incremental indexing fills in their local diff.

• Format: SQLite database, indexes stripped, VACUUM INTO compacted, then zstd 1.5.7 compressed (8–13:1 ratio typical)
• Two tiers:
  • Best (zstd -9 + index strip + VACUUM INTO) — written on explicit index_repository
  • Fast (zstd -3) — written by the watcher for low-latency incremental updates
• Bootstrap: when no local DB exists but the artifact is present, index_repository imports the artifact first, then runs incremental indexing — avoiding the full reindex cost
• No merge pain: a .codebase-memory/.gitattributes line with merge=ours is auto-created on first export, so concurrent edits don't produce conflicts on the binary artifact
• Commit it deliberately: the artifact is rewritten on every index, including the watcher's Fast tier, and git stores each rewrite as a full new blob. Committing every refresh is what turns a 20 MB file into gigabytes of history — one team reached ~6 GB across ~350 commits of this single path. Pick a cadence (a release, a milestone, a nightly job) rather than committing every save.
• Git LFS, if it must move on every commit: track it from the repo-root .gitattributes and leave the auto-created .codebase-memory/.gitattributes in place — the nearer file goes on supplying merge=ours, and only filter comes from the root:

  .codebase-memory/graph.db.zst filter=lfs diff=lfs merge=lfs -text
  

  Track only the .zst; artifact.json is small and carries the schema version. The attribute applies to future commits only, so a repo that already has the blobs in history needs git-filter-repo to rewrite them first. Two costs to weigh before adopting it: GitHub meters LFS storage and bandwidth, and its objects cannot be pruned without contacting support; and every teammate needs git lfs install — without it their checkout leaves a pointer file where the artifact should be, the integrity-checked import refuses it, and they fall back to a full reindex.
• Optional: never committed unless you want it. Add .codebase-memory/ to .gitignore if you prefer everyone to reindex from scratch.

The result is similar in spirit to graphify's graphify-out/ directory, but as a single compressed file with explicit two-tier export, integrity-checked import, and zero merge friction.

How It Works

codebase-memory-mcp is a structural analysis backend — it builds and queries the knowledge graph. It does not include an LLM. Instead, it relies on your MCP client (Claude Code, or any MCP-compatible agent) to be the intelligence layer.

You: "what calls ProcessOrder?"

Agent calls: trace_path(function_name="ProcessOrder", direction="inbound")

codebase-memory-mcp: executes graph query, returns structured results

Agent: presents the call chain in plain English

Why no built-in LLM? Other code graph tools embed an LLM for natural language → graph query translation. This means extra API keys, extra cost, and another model to configure. With MCP, the agent you're already talking to is the query translator.

Performance

Benchmarked on Apple M3 Pro:

RAM-first pipeline: All indexing runs in memory (LZ4 HC compressed read, in-memory SQLite, single dump at end). Memory is released back to the OS after indexing completes.

Token efficiency: Five structural queries consumed ~3,400 tokens via codebase-memory-mcp versus ~412,000 tokens via file-by-file grep exploration — a 99.2% reduction.

To measure comparable quality, latency, and agent-efficiency metrics on your own workload, see Measuring quality, latency, and agent savings. Exact reproduction of the figures above requires the original inputs and raw artifacts.

Troubleshooting & Diagnostics

codebase-memory-mcp runs 100% locally and collects no telemetry — your code, queries, environment, and usage never leave your machine. That privacy guarantee also means that when you hit something we can't reproduce on our side (a slow memory climb over hours, a performance regression, a leak that only appears after days of real use), we have no data at all unless you choose to send it. Here is how to capture it yourself.

Capture a diagnostics log

Set CBM_DIAGNOSTICS=1 before the first daemon-backed MCP session starts, then reproduce the problem (let it run as long as it takes — a slow leak needs time to show in the trend). The shared daemon captures this setting from the session that starts it. If it is already running, close all daemon-backed sessions so it exits before changing the setting. The daemon creates a fresh owner-private cbm-diagnostics-<pid>-<random> directory below the system temp directory ($TMPDIR or /tmp on macOS/Linux, %TEMP% on Windows). The exact paths are recorded by the diagnostics.start event in ${CBM_CACHE_DIR}/logs/cbm-daemon.log:

The private randomized directory prevents another local account from pre-placing a link or special file at a predictable diagnostics path. Its <pid> component is the shared daemon's process ID, also recorded by the daemon.start event. Set the variable consistently in the env block of each agent's MCP server config, or export it before launching the first session.

What to share

When you open a memory/performance issue, attach the .ndjson trajectory — it contains no source code or query text, only resource counters. If you'd rather not attach a file, paste it (or an agent's summary of it) into the issue: your assistant can read the NDJSON directly and report whether rss/committed grow monotonically, how fast, and relative to query count — which is exactly what we need to find the cause.

Installation

Pre-built Binaries

Every release includes checksums.txt with SHA-256 hashes. The executable is self-contained — no adjacent data file is required. Linux -portable archives contain the fully static builds; ordinary platform archives use their native system ABI.

Windows note: SmartScreen may show a warning for unsigned software. Click "More info" → "Run anyway". Verify integrity with checksums.txt.

Setup Scripts

macOS / Linux:

curl -fsSL https://raw.githubusercontent.com/DeusData/codebase-memory-mcp/main/scripts/setup.sh | bash

Windows (PowerShell):

irm https://raw.githubusercontent.com/DeusData/codebase-memory-mcp/main/scripts/setup-windows.ps1 | iex

AUR (Arch Linux)

yay -S codebase-memory-mcp-bin

paru -S codebase-memory-mcp-bin

The codebase-memory-mcp-bin package is available at: https://aur.archlinux.org/packages/codebase-memory-mcp-bin

Nix (flake)

The flake exposes two server packages plus the standalone frontend:

Run directly without installing:

# Standard server
nix run github:DeusData/codebase-memory-mcp

# Server with the embedded graph UI
nix run github:DeusData/codebase-memory-mcp#codebase-memory-mcp-ui -- --ui=true --port=9749
# then open http://127.0.0.1:9749

Or build a binary into ./result/bin/codebase-memory-mcp:

nix build github:DeusData/codebase-memory-mcp                          # standard
nix build github:DeusData/codebase-memory-mcp#codebase-memory-mcp-ui   # with UI

Working in a clone? Use . in place of the flake URL, e.g. nix run .#codebase-memory-mcp-ui -- --ui=true, or drop into a shell that puts the binary on PATH with nix shell .#codebase-memory-mcp-ui.

Note: launched by hand (not from an MCP client) the server exits as soon as stdin closes — that's normal MCP behaviour. Keep stdin open while testing the UI, e.g. sleep infinity | codebase-memory-mcp --ui=true --port=9749. The codebase-memory-mcp-ui package embeds the UI at build time; nix run'ing the standard default package with --ui=true will refuse to start the HTTP server.

Install via Claude Code

You: "Install this MCP server: https://github.com/DeusData/codebase-memory-mcp"

Build from Source

git clone https://github.com/DeusData/codebase-memory-mcp.git
cd codebase-memory-mcp
scripts/build.sh --with-ui          # the shipped composition (graph UI embedded)
scripts/build.sh                    # without the UI (development only)
# Binary at: build/c/codebase-memory-mcp   (codebase-memory-mcp.exe on Windows)

Every platform ships one self-contained executable: the graph UI and the agent integration templates are linked into the binary, so an extracted archive is immediately complete.

Run the test suite (6,768 tests across 120 suites):

scripts/test.sh                     # full: clean sanitizer build + all suites + guards
scripts/test.sh --suites <name>     # one suite, incremental, seconds
build/c/test-runner --list-suites   # what is available

scripts/test.sh is the same entry the CI gates run, so a local pass means the same thing a CI pass does. The canonical local artifact-flow check builds both stripped/unstripped candidates, defaults to the stripped candidate for this explicitly unscanned local run, packages those exact bytes, extracts the archive, and smokes it:

scripts/ci/smoke-artifact.sh <linux|darwin|windows> <amd64|arm64>

scripts/package-release.sh is intentionally a lower-level immutable boundary: it accepts only an already-final --selected-binary plus its --expected-sha256; it never builds, strips, signs, or relinks the executable.

Manual MCP Configuration

Add to ~/.claude.json (user scope) or project .mcp.json:

{
  "mcpServers": {
    "codebase-memory-mcp": {
      "command": "/path/to/codebase-memory-mcp",
      "args": []
    }
  }
}

Restart your agent. Verify with /mcp — you should see codebase-memory-mcp with 15 tools.

Multi-Agent Support

install configures 45 client surfaces: 39 detected automatically and 6 conditional or explicit. “Conditional” means the installer writes only when the documented platform or an explicit, already-existing config path proves the target is active. It never flips experimental feature flags, enables plugins, YOLO modes, global permission bypasses, or third-party instruction trust.

Where a client has a documented custom-agent format, the installer creates three exact-owned definitions from one canonical contract:

• Scout (Tier 1) — about 3–4 narrow calls for fast positive, provisional discovery; no absence, exhaustive-impact, or dead-code claims.
• Verify (Tier 2, default) — task-directed graph evidence, exact source checks, path coverage for every cited file, and scope coverage before negative claims.
• Auditor (Tier 3) — bounded scope, current index generation, complete relevant pagination, broader relationship checks, and explicit unresolved limitations.

Every direct tier batches check_index_coverage for its evidence paths and reads flagged ranges or skipped/excluded files directly. A clean coverage result means only “no recorded gap,” never proof of completeness. Clients without safe child MCP access receive the same three tiers as parent-handoff agents; the parent must supply project, generation, pagination state, graph evidence, and coverage results. Updates migrate only byte-identical prior Verify definitions and never overwrite user-modified agents.