Codevira

One memory layer for every AI coding tool you use. Switch between Claude Code, Cursor, Windsurf, and Antigravity without losing context, decisions, or progress.

Python License: MIT MCP Version PRs Welcome

Built for solo developers working on local projects with AI agents. Codevira gives every AI tool you use access to the same persistent project memory — so you stop re-explaining your codebase every session, stop losing carefully-made decisions, and stop burning tokens on re-discovery.

Works with: Claude Code · Claude Desktop · Cursor · Windsurf · Google Antigravity · any MCP-compatible AI tool

---

The Problem (Four Pains Codevira Solves)

If you code with AI agents on a project longer than a week, you've felt all of these:

1. Re-explaining your codebase every session

Every new chat starts from zero. The AI doesn't know your architecture, your conventions, your "we don't do it that way" decisions. You waste the first 10 minutes (and thousands of tokens) catching it up — only to do it again tomorrow.

2. AI undoing your careful decisions

Last week you debugged a tricky retry policy for 3 hours. Today's AI session refactors it to a simpler version because it has no idea why the complexity exists. Now it's broken again.

3. Cross-tool amnesia

You started planning in Claude Code. Switched to Cursor for autocomplete. Opened Antigravity to run tests. Three different agents, three different blind copies of your project state. Nothing carries over.

4. Token budget burned on re-discovery

Your AI agent reads the same 12 files every session before doing any actual work. You're paying API costs for the same lookups, over and over.

Codevira is a persistent memory layer that fixes all four — for every AI tool, on every project, on your local machine.

---

How It Works

Codevira is a Model Context Protocol server that runs locally and gives any AI tool a structured, queryable memory of your codebase:

Capability	What it means for you
Zero-config setup	pipx install codevira && codevira register — that's it. No prompts, no JSON editing. Auto-detects language, source dirs, and IDE configs
Cross-tool continuity	One get_session_context() call brings any AI agent up to speed in ~800 tokens — works identically in Claude Code, Cursor, Windsurf, Antigravity
Decision protection	do_not_revert flags + searchable decision log stop AI agents from undoing past architectural choices
Context graph	Every source file has a node: role, rules, dependencies, stability, blast radius. AI calls get_node(path) instead of re-reading the file
Function-level call graph	get_impact(file) answers "what breaks if I change this?" before the AI modifies anything
Semantic code search	Natural-language search across your codebase (search_codebase("auth flow"))
Roadmap + changesets	Multi-file work tracked atomically; sessions resume cleanly after interruption
Adaptive learning	Tracks which past decisions panned out — gives confidence scores and surfaces patterns
Cross-project memory	Learned preferences sync across all your local projects via ~/.codevira/global.db
Auto-init on first call	No codevira init needed — first MCP tool call triggers background project setup

Token-efficient by design

Codevira is built around the principle that AI agent context windows are precious. Tools return summaries by default with opt-in full data:

• get_node(path) — ~100 tokens by default (counts + flags). Pass full=true for the entire rules array.
• get_impact(path) — 10 affected files. Pass summary_only=true for just counts (~80 tokens) before deciding to dig deeper.
• search_codebase(query) — file/symbol pointers only. Pass include_content=true to inline source.
• search_decisions(query) — 5 truncated matches. Pass full=true for verbatim text.

The agent always asks for what it needs, in the size it needs.

---

Quick Start

1. Install

# Recommended: global install via pipx (isolated, works everywhere)
pipx install codevira

# Alternative: pip install
pip install codevira

Installs the full toolkit (23 AI-facing MCP tools + 12 admin/CLI tools) out of the box. Semantic search downloads a ~90MB embedding model on first use.

2. Register with your AI tools

codevira register

This one-time global command injects Codevira's MCP config into all detected AI tools — Claude Code, Cursor, Windsurf, Claude Desktop, and Google Antigravity. Run it from anywhere; no project directory needed.

3. Start using

Open any project in your AI tool. On the first MCP tool call, Codevira auto-initializes:

• Detects language, source directories, and file extensions from project markers
• Creates the context graph and roadmap
• Installs a post-commit git hook for automatic reindexing

No explicit codevira init needed — everything happens on demand.

Note: codevira init is still available for explicit per-project setup with custom settings.

4. Verify

Ask your AI agent to call get_roadmap() — it should return your current phase and next action.

Note: Restart your AI tool after running codevira register to pick up the new MCP config.

Manual config (only if auto-inject didn't detect your tool)

Codevira supports two transports. Use the right one for your client:

Client	Transport	Config file
Claude Desktop (app)	stdio	~/Library/Application Support/Claude/claude_desktop_config.json
Claude Code (CLI)	stdio or HTTP	.claude/settings.json
Cursor	stdio	.cursor/mcp.json
Windsurf	stdio	.windsurf/mcp.json
Google Antigravity	stdio	~/.gemini/antigravity/mcp_config.json

Stdio transport — Claude Desktop, Cursor, Windsurf (.claude/settings.json / .cursor/mcp.json / .windsurf/mcp.json):

{
  "mcpServers": {
    "codevira": {
      "command": "codevira",
      "args": [],
      "cwd": "/path/to/your-project"
    }
  }
}

Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "codevira": {
      "command": "/path/to/codevira",
      "args": ["--project-dir", "/path/to/your-project"]
    }
  }
}

Tip: find the full binary path with which codevira

HTTP/HTTPS transport — Preview in v1.7, single-project only. The HTTP server binds to one project at startup and cannot switch contexts per request. Multi-project HTTPS is planned for v1.8. For multi-project work today, use stdio via codevira register (above).

First start the HTTP server in a terminal:

codevira serve --port 7007 --project-dir /path/to/your-project
# For HTTPS (required by some clients):
codevira serve --https --port 7443 --project-dir /path/to/your-project

Then register the URL:

{
  "mcpServers": {
    "codevira": {
      "url": "https://localhost:7443/mcp"
    }
  }
}

HTTPS note: Claude Code uses Node.js, which requires a trusted CA for HTTPS. Run once to trust the mkcert CA:

brew install mkcert && mkcert -install
launchctl setenv NODE_EXTRA_CA_CERTS "$(mkcert -CAROOT)/rootCA.pem"
echo 'export NODE_EXTRA_CA_CERTS="$(mkcert -CAROOT)/rootCA.pem"' >> ~/.zshrc

Then restart Claude Code.

Auto-start on login (macOS):

codevira serve --install-service    # start server automatically on login
codevira serve --uninstall-service  # remove auto-start

Google Antigravity (~/.gemini/antigravity/mcp_config.json):

{
  "mcpServers": {
    "codevira": {
      "$typeName": "exa.cascade_plugins_pb.CascadePluginCommandTemplate",
      "command": "codevira",
      "args": []
    }
  }
}

Codevira data layout (v1.6)

~/.codevira/                         <- global Codevira home
├── global.db                        <- cross-project intelligence
├── projects/
│   └── <project-key>/               <- per-project data (keyed by path)
│       ├── config.yaml
│       ├── metadata.json
│       ├── graph/
│       │   ├── graph.db
│       │   └── changesets/
│       ├── codeindex/               <- semantic search (optional)
│       └── logs/
└── certs/                           <- HTTPS certs (if using --https)

Legacy .codevira/ directories inside project repos are auto-migrated to centralized storage on first server start.

Configuration

Each project has a config.yaml at ~/.codevira/projects/<project-key>/config.yaml. It's auto-generated on first use with sensible defaults, but you can edit it to customize what Codevira indexes:

project:
  name: my-project
  language: python
  collection_name: my_project
  # Which directories to scan for source files
  watched_dirs:
    - src
    - tests
    - scripts
  # Which file extensions count as "source" for indexing + change detection
  file_extensions:
    - .py
    - .ts
    - .tsx
  # Directories to skip even if inside watched_dirs
  skip_dirs:
    - node_modules
    - .venv
    - __pycache__
    - dist
    - build
logs:
  # 0 = keep sessions/decisions forever (default).
  # Only set > 0 if you have privacy reasons to time-bound history.
  retention_days: 0

Common gotchas:

• file_extensions must be a proper YAML list — each extension on its own line. This is wrong:

  file_extensions:
    - .py, .md, .html    # ❌ one item containing commas, not three extensions
  

  This is correct:

  file_extensions:
    - .py
    - .md
    - .html
  

  Or inline:

  file_extensions: [.py, .md, .html]
  

• file_extensions is intended for source code (Python, TypeScript, Go, Rust, etc.). Codevira uses tree-sitter AST parsing — putting .md or .html here may produce malformed graph nodes since tree-sitter parsers for those languages are different.

• Files are only scanned if they live inside watched_dirs. Adding an extension alone isn't enough — make sure the directory is listed too.

After editing the config, run codevira index --full to rebuild the graph from scratch, or codevira index for incremental changes.

Uninstall / Reset

codevira clean              # remove global data + IDE configs + launchd service
codevira clean --all        # also remove per-project artifacts
codevira clean --dry-run    # preview what would be removed

---

How It Works

Setup Flow

flowchart LR

A["pipx install codevira"] --> B["codevira register"]
B --> C["Open project in\nClaude Code / Cursor /\nWindsurf / Antigravity"]
C --> D["First MCP tool call\ntriggers auto-init"]
D --> E["✓ Config written\n✓ Graph built\n✓ Roadmap created\n✓ Ready"]

Agent Session Lifecycle

flowchart TB

Start([Start Session])

subgraph "Orientation (single call)"
A["get_session_context()\nroadmap + changesets +\ndecisions + global intelligence"]
end

subgraph "Work"
B[get_node / get_impact\nbefore touching files]
C[Plan + Implement + Test]
D[refresh_index\nafter changes]
end

subgraph "Session End"
E[update_node — record changes]
F[write_session_log — decisions]
G[update_next_action — handoff]
end

Start --> A
A --> B
B --> C
C --> D
D --> E
E --> F
F --> G

Architecture

flowchart TB

A[Source Code\n15+ languages]

subgraph "Indexing Pipeline"
B[Tree-sitter AST Parser]
C[Function / Class / Call Extraction]
D[Background File Watcher\nauto-reindex on save]
end

subgraph "Centralized Storage — ~/.codevira/"
E[(Context Graph + Call Graph\nSQLite DB)]
F[(Semantic Index\nChromaDB — optional)]
G[(Global Memory\ncross-project intelligence)]
H[(Session Logs + Decisions\nsearchable history)]
end

subgraph "Adaptive Learning"
I[Outcome Tracking]
J[Rule Inference]
K[Preference Learning]
end

subgraph "MCP Server"
L[36 Tools + 5 Prompts\nstdio or HTTP transport]
end

M[AI Coding Agent\nClaude Code · Cursor · Windsurf · Antigravity]

A --> B
B --> C
C --> E
C --> F
D --> B

E --> L
F --> L
G --> L
H --> L

I --> G
J --> G
K --> G

E --> I

L --> M

---

Session Protocol

Every agent session follows a simple protocol. Set it up once in your agent's system prompt — then your agents handle the rest.

Session start (mandatory):

list_open_changesets()      -> resume any unfinished work first
get_roadmap()               -> current phase, next action
search_decisions("topic")   -> check what's already been decided
get_node("src/service.py")  -> read rules before touching a file
get_impact("src/service.py") -> check blast radius

Session end (mandatory):

complete_changeset(id, decisions=[...])
update_node(file_path, changes)
update_next_action("what the next agent should do")
write_session_log(...)

This loop keeps every session fast, focused, and resumable.

---

MCP Tools + 5 Prompts

23 tools exposed to AI agents (token-optimized, summary-first). The remaining 12 tools are admin/dashboard tools that work via dispatch but aren't advertised in list_tools() — humans access them via the CLI or via specific MCP prompts. Tools marked (admin) below.

Graph Tools

Tool	Description
get_node(file_path, full?)	Summary by default (counts + flags); full=true for rules/dependencies arrays
get_impact(file_path, summary_only?)	Blast radius — summary_only=true returns just counts (~80 tokens)
update_node(file_path, changes)	Append rules, connections, key_functions
query_graph(file_path, symbol?, query_type)	Function-level: callers, callees, tests, dependents, symbols
list_nodes(...) (admin)	Bulk node listing — agents should use targeted queries instead
add_node(...) (admin)	Register a new file (auto-generated by refresh_graph)
refresh_graph(file_paths?) (admin)	Auto-generate stubs (background/automatic)
refresh_index(file_paths?) (admin)	Background reindex (fire-and-forget)
export_graph(format, scope?) (admin)	Mermaid/DOT export — large dump
get_graph_diff(base_ref?, head_ref?) (admin)	PR review — use review_changes prompt
analyze_changes(base_ref?, head_ref?) (admin)	Risk scoring — use pre_commit_check prompt
find_hotspots(threshold?) (admin)	Complexity dashboard

Roadmap Tools

Tool	Description
get_roadmap()	Current phase, next action, open changesets
get_phase(number)	Full details of any phase by number
update_next_action(text)	Set what the next agent should do
update_phase_status(status)	Mark phase in_progress / blocked
add_phase(phase, name, description, ...)	Queue new upcoming work
complete_phase(number, key_decisions)	Mark done, auto-advance to next
defer_phase(number, reason)	Move a phase to the deferred list
get_full_roadmap(include_decisions?) (admin)	Full history with all decisions inline

Changeset Tools

Tool	Description
list_open_changesets()	All in-progress changesets
start_changeset(id, description, files)	Open a multi-file changeset
complete_changeset(id, decisions)	Close and record decisions
update_changeset_progress(id, last_file, blocker?)	Mid-session checkpoint

Search Tools

Tool	Description
search_codebase(query, limit?, include_content?)	Semantic search — pointers only by default
search_decisions(query, limit?, full?)	Past decisions (default 5, truncated context)
get_history(file_path, limit?, full?)	Recent decisions touching a file (default 5)
write_session_log(...)	Write structured session record

Adaptive Learning Tools

Tool	Description
get_session_context()	THE main "catch me up" call — start every session here (~800 tokens)
get_decision_confidence(file_path?, pattern?)	Outcome-based reliability scores
get_preferences(category?) (admin)	Already in get_session_context
get_learned_rules(file_path?, category?) (admin)	Already in get_session_context
get_project_maturity() (admin)	Dashboard metric — use architecture_overview prompt

Code Reader Tools

Tool	Description
get_signature(file_path)	All public symbols, signatures, line numbers (Python, TypeScript, Go, Rust)
get_code(file_path, symbol)	Full source of one function or class

Playbook Tool

Tool	Description
get_playbook(task_type)	Curated rules for: add_tool, add_service, add_schema, debug_pipeline, commit, write_test

MCP Workflow Prompts (v1.5)

Prompt	Description
review_changes	Staged diff + blast radius + risk score
debug_issue	Symptom -> affected files -> call chain -> hypothesis
onboard_session	Full project context catch-up for new sessions
pre_commit_check	Test coverage gaps + high-risk functions before commit
architecture_overview	Module map + hotspots + dependency summary

---

Language Support

Feature	Python	TypeScript	Go	Rust	12+ Others
Context graph + blast radius	Y	Y	Y	Y	Y
Semantic code search	Y	Y	Y	Y	Y
Function-level call graph	Y	Y	Y	Y
get_signature / get_code	Y	Y	Y	Y
AST-based chunking	Y	Y	Y	Y
Auto-generated graph stubs	Y	Y	Y	Y
Roadmap + changesets	Y	Y	Y	Y	Y
Session logs + decision search	Y	Y	Y	Y	Y

Supported languages: Python, TypeScript, JavaScript, Go, Rust, Java, Kotlin, C#, Ruby, PHP, C, C++, Swift, Solidity, Vue.

---

Requirements

• Python 3.10+
• ~500MB install (includes ChromaDB + sentence-transformers for semantic search)
• ~90MB model download on first search_codebase() call

pip install codevira includes the full toolkit out of the box — graph, roadmap, changesets, code reader, learning, call graph, and semantic search.

Minimal install (no semantic search)

If you want to skip the ML stack and use only graph-based tools (semantic search disabled), install without the search deps:

pip install codevira --no-deps
pip install pyyaml mcp watchdog tree-sitter tree-sitter-language-pack rich uvicorn starlette pathspec

The search_codebase tool will be hidden from your AI agent; all other tools work normally.

---

Background

Want to understand the full story behind why this was built, the design decisions, what didn't work, and how it compares to other tools in the ecosystem?

Read the full write-up: How We Cut AI Coding Agent Token Usage by 92%

---

Contributing

Contributions are welcome. Read CONTRIBUTING.md for the full guide.

Reporting a bug? Open a bug report Requesting a feature? Open a feature request Found a security issue? Read SECURITY.md — please don't use public issues for vulnerabilities.

---

FAQ

Common questions about setup, usage, architecture, and troubleshooting — see FAQ.md.

Roadmap

See what's built, what's next, and the long-term vision — see ROADMAP.md.

License

MIT — free to use, modify, and distribute.