Allows connecting to ChatGPT via mcpo to enable tool calling for code editing and retrieval, leveraging Serena's semantic code understanding.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@Serena MCP Server Find all references to the 'Config' class"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
:rocket: Serena is a powerful coding agent toolkit capable of turning an LLM into a fully-featured agent that works directly on your codebase. Unlike most other tools, it is not tied to an LLM, framework or an interface, making it easy to use it in a variety of ways.
:wrench: Serena provides essential semantic code retrieval and editing tools that are akin to an IDE's capabilities, extracting code entities at the symbol level and exploiting relational structure. When combined with an existing coding agent, these tools greatly enhance (token) efficiency.
:free: Serena is free & open-source, enhancing the capabilities of LLMs you already have access to free of charge.
You can think of Serena as providing IDE-like tools to your LLM/coding agent. With it, the agent no longer needs to read entire
files, perform grep-like searches or string replacements to find and edit the right code. Instead, it can use code centered tools like find_symbol, find_referencing_symbols and insert_after_symbol.
Serena provides the necessary tools for coding workflows, but an LLM is required to do the actual work, orchestrating tool use.
For example, supercharge the performance of Claude Code with a one-line shell command.
In general, Serena can be integrated with an LLM in several ways:
by using the model context protocol (MCP). Serena provides an MCP server which integrates with
by using mcpo to connect it to ChatGPT or other clients that don't support MCP but do support tool calling via OpenAPI.
by incorporating Serena's tools into an agent framework of your choice, as illustrated here. Serena's tool implementation is decoupled from the framework-specific code and can thus easily be adapted to any agent framework.
A demonstration of Serena efficiently retrieving and editing code within Claude Code, thereby saving tokens and time. Efficient operations are not only useful for saving costs, but also for generally improving the generated code's quality. This effect may be less pronounced in very small projects, but often becomes of crucial importance in larger ones.
https://github.com/user-attachments/assets/ab78ebe0-f77d-43cc-879a-cc399efefd87
A demonstration of Serena implementing a small feature for itself (a better log GUI) with Claude Desktop. Note how Serena's tools enable Claude to find and edit the right symbols.
https://github.com/user-attachments/assets/6eaa9aa1-610d-4723-a2d6-bf1e487ba753
Serena's semantic code analysis capabilities build on language servers using the widely implemented language server protocol (LSP). The LSP provides a set of versatile code querying and editing functionalities based on symbolic understanding of the code. Equipped with these capabilities, Serena discovers and edits code just like a seasoned developer making use of an IDE's capabilities would. Serena can efficiently find the right context and do the right thing even in very large and complex projects! So not only is it free and open-source, it frequently achieves better results than existing solutions that charge a premium.
Language servers provide support for a wide range of programming languages. With Serena, we provide direct, out-of-the-box support for:
Python
TypeScript/Javascript
PHP (uses Intelephense LSP; set INTELEPHENSE_LICENSE_KEY environment variable for premium features)
Go (requires installation of gopls)
R (requires installation of the languageserver R package)
Rust (requires rustup - uses rust-analyzer from your toolchain)
C/C++ (you may experience issues with finding references, we are working on it)
Zig (requires installation of ZLS - Zig Language Server)
C#
Ruby (by default, uses ruby-lsp, specify ruby_solargraph as your language to use the previous solargraph based implementation)
Swift
Kotlin (uses the pre-alpha official kotlin LS, some issues may appear)
Java (Note: startup is slow, initial startup especially so. There may be issues with java on macos and linux, we are working on it.)
Clojure
Dart
Bash
Lua (automatically downloads lua-language-server if not installed)
Nix (requires nixd installation)
Elixir (requires installation of NextLS and Elixir; Windows not supported)
Erlang (requires installation of beam and erlang_ls, experimental, might be slow or hang)
Support for further languages can easily be added by providing a shallow adapter for a new language server implementation, see Serena's memory on that.
Most users report that Serena has strong positive effects on the results of their coding agents, even when used within very capable agents like Claude Code. Serena is often described to be a game changer, providing an enormous productivity boost.
Serena excels at navigating and manipulating complex codebases, providing tools that support precise code retrieval and editing in the presence of large, strongly structured codebases. However, when dealing with tasks that involve only very few/small files, you may not benefit from including Serena on top of your existing coding agent. In particular, when writing code from scratch, Serena will not provide much value initially, as the more complex structures that Serena handles more gracefully than simplistic, file-based approaches are yet to be created.
Several videos and blog posts have talked about Serena:
YouTube:
Blog posts:
Related MCP server: LSP-MCP
Serena can be used in various ways, below you will find instructions for selected integrations.
For coding with Claude, we recommend using Serena through Claude Code or Claude Desktop. You can also use Serena in most other terminal-based clients.
If you want a GUI experience outside an IDE, you can use one of the many local GUIs that support MCP servers. You can also connect Serena to many web clients (including ChatGPT) using mcpo.
If you want to use Serena integrated in your IDE, see the section on other MCP clients.
You can use Serena as a library for building your own applications. We try to keep the public API stable, but you should still expect breaking changes and pin Serena to a fixed version if you use it as a dependency.
Serena is managed by uv, so you will need to install it).
You have several options for running the MCP server, which are explained in the subsections below.
The typical usage involves the client (Claude Code, Claude Desktop, etc.) running the MCP server as a subprocess (using stdio communication), so the client needs to be provided with the command to run the MCP server. (Alternatively, you can run the MCP server in SSE mode and tell your client how to connect to it.)
Note that no matter how you run the MCP server, Serena will, by default, start a small web-based dashboard on localhost that will display logs and allow shutting down the MCP server (since many clients fail to clean up processes correctly). This and other settings can be adjusted in the configuration and/or by providing command-line arguments.
uvx can be used to run the latest version of Serena directly from the repository, without an explicit local installation.
uvx --from git+https://github.com/oraios/serena serena start-mcp-serverExplore the CLI to see some of the customization options that serena provides (more info on them below).
Clone the repository and change into it.
git clone https://github.com/oraios/serena
cd serenaOptionally edit the configuration file in your home directory with
uv run serena config editIf you just want the default config, you can skip this part, and a config file will be created when you first run Serena.
Run the server with uv:
uv run serena start-mcp-serverWhen running from outside the serena installation directory, be sure to pass it, i.e., use
uv run --directory /abs/path/to/serena serena start-mcp-server⚠️ Docker support is currently experimental with several limitations. Please read the Docker documentation for important caveats before using it.
You can run the Serena MCP server directly via docker as follows,
assuming that the projects you want to work on are all located in /path/to/your/projects:
docker run --rm -i --network host -v /path/to/your/projects:/workspaces/projects ghcr.io/oraios/serena:latest serena start-mcp-server --transport stdioReplace /path/to/your/projects with the absolute path to your projects directory. The Docker approach provides:
Better security isolation for shell command execution
No need to install language servers and dependencies locally
Consistent environment across different systems
Alternatively, use docker compose with the compose.yml file provided in the repository.
See the Docker documentation for detailed setup instructions, configuration options, and known limitations.
If you are using Nix and have enabled the nix-command and flakes features, you can run Serena using the following command:
nix run github:oraios/serena -- start-mcp-server --transport stdioYou can also install Serena by referencing this repo (github:oraios/serena) and using it in your Nix flake. The package is exported as serena.
ℹ️ Note that MCP servers which use stdio as a protocol are somewhat unusual as far as client/server architectures go, as the server necessarily has to be started by the client in order for communication to take place via the server's standard input/output stream. In other words, you do not need to start the server yourself. The client application (e.g. Claude Desktop) takes care of this and therefore needs to be configured with a launch command.
When using instead the SSE mode, which uses HTTP-based communication, you control the server lifecycle yourself, i.e. you start the server and provide the client with the URL to connect to it.
Simply provide start-mcp-server with the --transport sse option and optionally provide the port.
For example, to run the Serena MCP server in SSE mode on port 9121 using a local installation,
you would run this command from the Serena directory,
uv run serena start-mcp-server --transport sse --port 9121and then configure your client to connect to http://localhost:9121/sse.
The Serena MCP server supports a wide range of additional command-line options, including the option to run in SSE mode and to adapt Serena to various contexts and modes of operation.
Run with parameter --help to get a list of available options.
Serena is very flexible in terms of configuration. While for most users, the default configurations will work,
you can fully adjust it to your needs by editing a few yaml files. You can disable tools, change Serena's instructions
(what we denote as the system_prompt), adjust the output of tools that just provide a prompt, and even adjust tool descriptions.
Serena is configured in four places:
The serena_config.yml for general settings that apply to all clients and projects.
It is located in your user directory under .serena/serena_config.yml.
If you do not explicitly create the file, it will be auto-generated when you first run Serena.
You can edit it directly or use
uvx --from git+https://github.com/oraios/serena serena config edit(or use the --directory command version).
In the arguments passed to the start-mcp-server in your client's config (see below),
which will apply to all sessions started by the respective client. In particular, the context parameter
should be set appropriately for Serena to be best adjusted to existing tools and capabilities of your client.
See for a detailed explanation. You can override all entries from the serena_config.yml through command line arguments.
In the .serena/project.yml file within your project. This will hold project-level configuration that is used whenever
that project is activated. This file will be autogenerated when you first use Serena on that project, but you can also
generate it explicitly with
uvx --from git+https://github.com/oraios/serena serena project generate-yml(or use the --directory command version).
Through the context and modes. Explore the modes and contexts section for more details.
After the initial setup, continue with one of the sections below, depending on how you want to use Serena.
If you are mostly working with the same project, you can configure to always activate it at startup
by passing --project <path_or_name> to the start-mcp-server command in your client's MCP config.
This is especially useful for clients which configure MCP servers on a per-project basis, like Claude Code.
Otherwise, the recommended way is to just ask the LLM to activate a project by providing it an absolute path to, or, in case the project was activated in the past, by its name. The default project name is the directory name.
"Activate the project /path/to/my_project"
"Activate the project my_project"
All projects that have been activated will be automatically added to your serena_config.yml, and for each
project, the file .serena/project.yml will be generated. You can adjust the latter, e.g., by changing the name
(which you refer to during the activation) or other options. Make sure to not have two different projects with the
same name.
ℹ️ For larger projects, we recommend that you index your project to accelerate Serena's tools; otherwise the first tool application may be very slow. To do so, run this from the project directory (or pass the path to the project as an argument):
uvx --from git+https://github.com/oraios/serena serena project index(or use the --directory command version).
Serena is a great way to make Claude Code both cheaper and more powerful!
From your project directory, add serena with a command like this,
claude mcp add serena -- <serena-mcp-server> --context ide-assistant --project $(pwd)where <serena-mcp-server> is your way of running the Serena MCP server.
For example, when using uvx, you would run
claude mcp add serena -- uvx --from git+https://github.com/oraios/serena serena start-mcp-server --context ide-assistant --project $(pwd)ℹ️ Serena comes with an instruction text, and Claude needs to read it to properly use Serena's tools.
As of version v1.0.52, claude code reads the instructions of the MCP server, so this is handled automatically.
If you are using an older version, or if Claude fails to read the instructions, you can ask it explicitly
to "read Serena's initial instructions" or run /mcp__serena__initial_instructions to load the instruction text.
If you want to make use of that, you will have to enable the corresponding tool explicitly by adding initial_instructions to the included_optional_tools
in your config.
Note that you may have to make Claude read the instructions when you start a new conversation and after any compacting operation to ensure Claude remains properly configured to use Serena's tools.
Serena works with OpenAI's Codex CLI out of the box, but you have to use the codex context for it to work properly. (The technical reason is that Codex doesn't fully support the MCP specifications, so some massaging of tools is required.).
Unlike Claude Code, in Codex you add an MCP server globally and not per project. Add the following to
~/.codex/config.toml (create the file if it does not exist):
[mcp_servers.serena]
command = "uvx"
args = ["--from", "git+https://github.com/oraios/serena", "serena", "start-mcp-server", "--context", "codex"]After codex has started, you need to activate the project, which you can do by saying:
"Activate the current dir as project using serena"
If you don't activate the project, you will not be able to use Serena's tools!
That's it! Have a look at ~/.codex/log/codex-tui.log to see if any errors occurred.
The Serena dashboard will run if you have not disabled it in the configuration, but due to Codex's sandboxing the webbrowser
may not open automatically. You can open it manually by going to http://localhost:24282/dashboard/index.html (or a higher port, if
that was already taken).
Codex will often show the tools as
failedeven though they are successfully executed. This is not a problem, seems to be a bug in Codex. Despite the error message, everything works as expected.
There are many terminal-based coding assistants that support MCP servers, such as Codex, Gemini-CLI, Qwen3-Coder, rovodev, the OpenHands CLI and opencode.
They generally benefit from the symbolic tools provided by Serena. You might want to customize some aspects of Serena by writing your own context, modes or prompts to adjust it to your workflow, to other MCP servers you are using, and to the client's internal capabilities.
For Claude Desktop (available for Windows and macOS), go to File / Settings / Developer / MCP Servers / Edit Config,
which will let you open the JSON file claude_desktop_config.json.
Add the serena MCP server configuration, using a run command depending on your setup.
local installation:
{
"mcpServers": {
"serena": {
"command": "/abs/path/to/uv",
"args": ["run", "--directory", "/abs/path/to/serena", "serena", "start-mcp-server"]
}
}
}uvx:
{
"mcpServers": {
"serena": {
"command": "/abs/path/to/uvx",
"args": ["--from", "git+https://github.com/oraios/serena", "serena", "start-mcp-server"]
}
}
}docker:
{
"mcpServers": {
"serena": {
"command": "docker",
"args": ["run", "--rm", "-i", "--network", "host", "-v", "/path/to/your/projects:/workspaces/projects", "ghcr.io/oraios/serena:latest", "serena", "start-mcp-server", "--transport", "stdio"]
}
}
}If you are using paths containing backslashes for paths on Windows
(note that you can also just use forward slashes), be sure to escape them correctly (\\).
That's it! Save the config and then restart Claude Desktop. You are ready for activating your first project.
ℹ️ You can further customize the run command using additional arguments (see above).
Note: on Windows and macOS there are official Claude Desktop applications by Anthropic, for Linux there is an open-source community version.
⚠️ Be sure to fully quit the Claude Desktop application, as closing Claude will just minimize it to the system tray – at least on Windows.
⚠️ Some clients may leave behind zombie processes. You will have to find and terminate them manually then. With Serena, you can activate the dashboard to prevent unnoted processes and also use the dashboard for shutting down Serena.
After restarting, you should see Serena's tools in your chat interface (notice the small hammer icon).
For more information on MCP servers with Claude Desktop, see the official quick start guide.
Being an MCP Server, Serena can be included in any MCP Client. The same configuration as above,
perhaps with small client-specific modifications, should work. Most of the popular
existing coding assistants (IDE extensions or VSCode-like IDEs) support connections
to MCP Servers. It is recommended to use the ide-assistant context for these integrations by adding "--context", "ide-assistant" to the args in your MCP client's configuration. Including Serena generally boosts their performance
by providing them tools for symbolic operations.
In this case, the billing for the usage continues to be controlled by the client of your choice (unlike with the Claude Desktop client). But you may still want to use Serena through such an approach, e.g., for one of the following reasons:
You are already using a coding assistant (say Cline or Cursor) and just want to make it more powerful.
You are on Linux and don't want to use the community-created Claude Desktop.
You want tighter integration of Serena into your IDE and don't mind paying for that.
Over the last months, several technologies have emerged that allow you to run a powerful local GUI and connect it to an MCP server. They will work with Serena out of the box. Some of the leading open source GUI technologies offering this are Jan, OpenHands, OpenWebUI and Agno. They allow combining Serena with almost any LLM (including locally running ones) and offer various other integrations.
Serena combines tools for semantic code retrieval with editing capabilities and shell execution. Serena's behavior can be further customized through Modes and Contexts. Find the complete list of tools below.
The use of all tools is generally recommended, as this allows Serena to provide the most value: Only by executing shell commands (in particular, tests) can Serena identify and correct mistakes autonomously.
However, it should be noted that the execute_shell_command tool allows for arbitrary code execution.
When using Serena as an MCP Server, clients will typically ask the user for permission
before executing a tool, so as long as the user inspects execution parameters beforehand,
this should not be a problem.
However, if you have concerns, you can choose to disable certain commands in your project's
.yml configuration file.
If you only want to use Serena purely for analyzing code and suggesting implementations
without modifying the codebase, you can enable read-only mode by setting read_only: true in your project configuration file.
This will automatically disable all editing tools and prevent any modifications to your codebase while still
allowing all analysis and exploration capabilities.
In general, be sure to back up your work and use a version control system in order to avoid losing any work.
Serena's behavior and toolset can be adjusted using contexts and modes. These allow for a high degree of customization to best suit your workflow and the environment Serena is operating in.
A context defines the general environment in which Serena is operating. It influences the initial system prompt and the set of available tools. A context is set at startup when launching Serena (e.g., via CLI options for an MCP server or in the agent script) and cannot be changed during an active session.
Serena comes with pre-defined contexts:
desktop-app: Tailored for use with desktop applications like Claude Desktop. This is the default.
agent: Designed for scenarios where Serena acts as a more autonomous agent, for example, when used with Agno.
ide-assistant: Optimized for integration into IDEs like VSCode, Cursor, or Cline, focusing on in-editor coding assistance.
Choose the context that best matches the type of integration you are using.
When launching Serena, specify the context using --context <context-name>.
Note that for cases where parameter lists are specified (e.g. Claude Desktop), you must add two parameters to the list.
If you are using a local server (such as Llama.cpp) which requires you to use OpenAI-compatible tool descriptions, use context oaicompat-agent instead of agent.
Modes further refine Serena's behavior for specific types of tasks or interaction styles. Multiple modes can be active simultaneously, allowing you to combine their effects. Modes influence the system prompt and can also alter the set of available tools by excluding certain ones.
Examples of built-in modes include:
planning: Focuses Serena on planning and analysis tasks.
editing: Optimizes Serena for direct code modification tasks.
interactive: Suitable for a conversational, back-and-forth interaction style.
one-shot: Configures Serena for tasks that should be completed in a single response, often used with planning for generating reports or initial plans.
no-onboarding: Skips the initial onboarding process if it's not needed for a particular session.
onboarding: (Usually triggered automatically) Focuses on the project onboarding process.
Modes can be set at startup (similar to contexts) but can also be switched dynamically during a session. You can instruct the LLM to use the switch_modes tool to activate a different set of modes (e.g., "switch to planning and one-shot modes").
When launching Serena, specify modes using --mode <mode-name>; multiple modes can be specified, e.g. --mode planning --mode no-onboarding.
:warning: Mode Compatibility: While you can combine modes, some may be semantically incompatible (e.g., interactive and one-shot). Serena currently does not prevent incompatible combinations; it is up to the user to choose sensible mode configurations.
You can create your own contexts and modes to precisely tailor Serena to your needs in two ways:
You can use Serena's CLI to manage modes and contexts. Check out
uvx --from git+https://github.com/oraios/serena serena mode --helpand
uvx --from git+https://github.com/oraios/serena serena context --help NOTE: Custom contexts/modes are simply YAML files in <home>/.serena, they are automatically registered and available for use by their name (filename without the .yml extension). If you don't want to use Serena's CLI, you can create and manage them in any way you see fit.
Using external YAML files: When starting Serena, you can also provide an absolute path to a custom .yml file for a context or mode.
This customization allows for deep integration and adaptation of Serena to specific project requirements or personal preferences.
By default, Serena will perform an onboarding process when it is started for the first time for a project. The goal of the onboarding is for Serena to get familiar with the project and to store memories, which it can then draw upon in future interactions. If an LLM should fail to complete the onboarding and does not actually write the respective memories to disk, you may need to ask it to do so explicitly.
The onboarding will usually read a lot of content from the project, thus filling up the context. It can therefore be advisable to switch to another conversation once the onboarding is complete. After the onboarding, we recommend that you have a quick look at the memories and, if necessary, edit them or add additional ones.
Memories are files stored in .serena/memories/ in the project directory,
which the agent can choose to read in subsequent interactions.
Feel free to read and adjust them as needed; you can also add new ones manually.
Every file in the .serena/memories/ directory is a memory file.
Whenever Serena starts working on a project, the list of memories is
provided, and the agent can decide to read them.
We found that memories can significantly improve the user experience with Serena.
Serena uses the code structure for finding, reading and editing code. This means that it will work well with well-structured code but may perform poorly on fully unstructured one (like a "God class" with enormous, non-modular functions). Furthermore, for languages that are not statically typed, type annotations are highly beneficial.
It is best to start a code generation task from a clean git state. Not only will
this make it easier for you to inspect the changes, but also the model itself will
have a chance of seeing what it has changed by calling git diff and thereby
correct itself or continue working in a followup conversation if needed.
:warning: Important: since Serena will write to files using the system-native line endings
and it might want to look at the git diff, it is important to
set git config core.autocrlf to true on Windows.
With git config core.autocrlf set to false on Windows, you may end up with huge diffs
only due to line endings. It is generally a good idea to globally enable this git setting on Windows:
git config --global core.autocrlf trueSerena can successfully complete tasks in an agent loop, where it iteratively acquires information, performs actions, and reflects on the results. However, Serena cannot use a debugger; it must rely on the results of program executions, linting results, and test results to assess the correctness of its actions. Therefore, software that is designed to meaningful interpretable outputs (e.g. log messages) and that has a good test coverage is much easier to work with for Serena.
We generally recommend to start an editing task from a state where all linting checks and tests pass.
We found that it is often a good idea to spend some time conceptualizing and planning a task before actually implementing it, especially for non-trivial task. This helps both in achieving better results and in increasing the feeling of control and staying in the loop. You can make a detailed plan in one session, where Serena may read a lot of your code to build up the context, and then continue with the implementation in another (potentially after creating suitable memories).
In our experience, LLMs are bad at counting, i.e. they have problems inserting blocks of code in the right place. Most editing operations can be performed at the symbolic level, allowing this problem is overcome. However, sometimes, line-level insertions are useful.
Serena is instructed to double-check the line numbers and any code blocks that it will edit, but you may find it useful to explicitly tell it how to edit code if you run into problems. We are working on making Serena's editing capabilities more robust.
For long and complicated tasks, or tasks where Serena has read a lot of content, you may come close to the limits of context tokens. In that case, it is often a good idea to continue in a new conversation. Serena has a dedicated tool to create a summary of the current state of the progress and all relevant info for continuing it. You can request to create this summary and write it to a memory. Then, in a new conversation, you can just ask Serena to read the memory and continue with the task. In our experience, this worked really well. On the up-side, since in a single session there is no summarization involved, Serena does not usually get lost (unlike some other agents that summarize under the hood), and it is also instructed to occasionally check whether it's on the right track.
Moreover, Serena is instructed to be frugal with context (e.g., to not read bodies of code symbols unnecessarily), but we found that Claude is not always very good in being frugal (Gemini seemed better at it). You can explicitly instruct it to not read the bodies if you know that it's not needed.
When using Serena through an MCP Client, you can use it together with other MCP servers. However, beware of tool name collisions! See info on that above.
Currently, there is a collision with the popular Filesystem MCP Server. Since Serena also provides filesystem operations, there is likely no need to ever enable these two simultaneously.
Serena provides two convenient ways of accessing the logs of the current session:
via the web-based dashboard (enabled by default)
This is supported on all platforms.
By default, it will be accessible at http://localhost:24282/dashboard/index.html,
but a higher port may be used if the default port is unavailable/multiple instances are running.
via the GUI tool (disabled by default)
This is mainly supported on Windows, but it may also work on Linux; macOS is unsupported.
Both can be enabled, configured or disabled in Serena's configuration file (serena_config.yml, see above).
If enabled, they will automatically be opened as soon as the Serena agent/MCP server is started.
The web dashboard will display usage statistics of Serena's tools if you set record_tool_usage_stats: True in your config.
In addition to viewing logs, both tools allow to shut down the Serena agent. This function is provided, because clients like Claude Desktop may fail to terminate the MCP server subprocess when they themselves are closed.
Support for MCP Servers in Claude Desktop and the various MCP Server SDKs are relatively new developments and may display instabilities.
The working configuration of an MCP server may vary from platform to platform and from client to client. We recommend always using absolute paths, as relative paths may be sources of errors. The language server is running in a separate sub-process and is called with asyncio – sometimes a client may make it crash. If you have Serena's log window enabled, and it disappears, you'll know what happened.
Some clients may not properly terminate MCP servers, look out for hanging python processes and terminate them manually, if needed.
To our knowledge, Serena is the first fully-featured coding agent where the entire functionality is available through an MCP server, thus not requiring API keys or subscriptions.
Many prominent subscription-based coding agents are parts of IDEs like Windsurf, Cursor and VSCode. Serena's functionality is similar to Cursor's Agent, Windsurf's Cascade or VSCode's agent mode.
Serena has the advantage of not requiring a subscription. A potential disadvantage is that it is not directly integrated into an IDE, so the inspection of newly written code is not as seamless.
More technical differences are:
Serena is not bound to a specific IDE or CLI. Serena's MCP server can be used with any MCP client (including some IDEs), and the Agno-based agent provides additional ways of applying its functionality.
Serena is not bound to a specific large language model or API.
Serena navigates and edits code using a language server, so it has a symbolic understanding of the code. IDE-based tools often use a RAG-based or purely text-based approach, which is often less powerful, especially for large codebases.
Serena is open-source and has a small codebase, so it can be easily extended and modified.
An alternative to subscription-based agents are API-based agents like Claude Code, Cline, Aider, Roo Code and others, where the usage costs map directly to the API costs of the underlying LLM. Some of them (like Cline) can even be included in IDEs as an extension. They are often very powerful and their main downside are the (potentially very high) API costs.
Serena itself can be used as an API-based agent (see the section on Agno above). We have not yet written a CLI tool or a dedicated IDE extension for Serena (and there is probably no need for the latter, as Serena can already be used with any IDE that supports MCP servers). If there is demand for a Serena as a CLI tool like Claude Code, we will consider writing one.
The main difference between Serena and other API-based agents is that Serena can also be used as an MCP server, thus not requiring an API key and bypassing the API costs. This is a unique feature of Serena.
There are other MCP servers designed for coding, like DesktopCommander and codemcp. However, to the best of our knowledge, none of them provide semantic code retrieval and editing tools; they rely purely on text-based analysis. It is the integration of language servers and the MCP that makes Serena unique and so powerful for challenging coding tasks, especially in the context of larger codebases.
We built Serena on top of multiple existing open-source technologies, the most important ones being:
multilspy. A library which wraps language server implementations and adapts them for interaction via Python and which provided the basis for our library Solid-LSP (src/solidlsp). Solid-LSP provides pure synchronous LSP calls and extends the original library with the symbolic logic that Serena required.
Agno and the associated agent-ui, which we use to allow Serena to work with any model, beyond the ones supporting the MCP.
All the language servers that we use through Solid-LSP.
Without these projects, Serena would not have been possible (or would have been significantly more difficult to build).
It is straightforward to extend Serena's AI functionality with your own ideas.
Simply implement a new tool by subclassing
serena.agent.Tool and implement the apply method with a signature
that matches the tool's requirements.
Once implemented, SerenaAgent will automatically have access to the new tool.
It is also relatively straightforward to add support for a new programming language.
We look forward to seeing what the community will come up with! For details on contributing, see contributing guidelines.
Here is the list of Serena's default tools with a short description (output of uv run serena tools list):
activate_project: Activates a project by name.
check_onboarding_performed: Checks whether project onboarding was already performed.
create_text_file: Creates/overwrites a file in the project directory.
delete_memory: Deletes a memory from Serena's project-specific memory store.
execute_shell_command: Executes a shell command.
find_file: Finds files in the given relative paths
find_referencing_symbols: Finds symbols that reference the symbol at the given location (optionally filtered by type).
find_symbol: Performs a global (or local) search for symbols with/containing a given name/substring (optionally filtered by type).
get_symbols_overview: Gets an overview of the top-level symbols defined in a given file.
insert_after_symbol: Inserts content after the end of the definition of a given symbol.
insert_before_symbol: Inserts content before the beginning of the definition of a given symbol.
list_dir: Lists files and directories in the given directory (optionally with recursion).
list_memories: Lists memories in Serena's project-specific memory store.
onboarding: Performs onboarding (identifying the project structure and essential tasks, e.g. for testing or building).
prepare_for_new_conversation: Provides instructions for preparing for a new conversation (in order to continue with the necessary context).
read_file: Reads a file within the project directory.
read_memory: Reads the memory with the given name from Serena's project-specific memory store.
replace_regex: Replaces content in a file by using regular expressions.
replace_symbol_body: Replaces the full definition of a symbol.
search_for_pattern: Performs a search for a pattern in the project.
think_about_collected_information: Thinking tool for pondering the completeness of collected information.
think_about_task_adherence: Thinking tool for determining whether the agent is still on track with the current task.
think_about_whether_you_are_done: Thinking tool for determining whether the task is truly completed.
write_memory: Writes a named memory (for future reference) to Serena's project-specific memory store.
There are several tools that are disabled by default, and have to be enabled explicitly, e.g., through the context or modes.
Note that several of our default contexts do enable some of these tools. For example, the desktop-app context enables the execute_shell_command tool.
The full list of optional tools is (output of uv run serena tools list --only-optional):
delete_lines: Deletes a range of lines within a file.
get_current_config: Prints the current configuration of the agent, including the active and available projects, tools, contexts, and modes.
initial_instructions: Gets the initial instructions for the current project.
Should only be used in settings where the system prompt cannot be set,
e.g. in clients you have no control over, like Claude Desktop.
insert_at_line: Inserts content at a given line in a file.
jet_brains_find_referencing_symbols: Finds symbols that reference the given symbol
jet_brains_find_symbol: Performs a global (or local) search for symbols with/containing a given name/substring (optionally filtered by type).
jet_brains_get_symbols_overview: Retrieves an overview of the top-level symbols within a specified file
remove_project: Removes a project from the Serena configuration.
replace_lines: Replaces a range of lines within a file with new content.
restart_language_server: Restarts the language server, may be necessary when edits not through Serena happen.
summarize_changes: Provides instructions for summarizing the changes made to the codebase.
switch_modes: Activates modes by providing a list of their names
Activates the project with the given name.
Inputs: project (required). Returns: result
Checks whether project onboarding was already performed. You should always call this tool before beginning to actually work on the project/after activating a project, but after calling the initial instructions tool.
Inputs: no named parameters. Returns: result
Write a new file or overwrite an existing file. Returns a message indicating success or failure.
Inputs: content (required), relative_path (required). Returns: result
Delete a memory file. Should only happen if a user asks for it explicitly, for example by saying that the information retrieved from a memory file is no longer correct or no longer relevant for the project.
Inputs: memory_file_name (required). Returns: result
Execute a shell command and return its output. If there is a memory about suggested commands, read that first.
Never execute unsafe shell commands like rm -rf / or similar!. Returns a JSON object containing the command's stdout and optionally stderr output.
Inputs: command (required), cwd, capture_stderr, max_answer_chars. Returns: result
Finds non-gitignored files matching the given file mask within the given relative path. Returns a JSON object with the list of matching files.
Inputs: file_mask (required), relative_path (required). Returns: result
Finds references to the symbol at the given name_path. The result will contain metadata about the referencing symbols
as well as a short code snippet around the reference. Returns a list of JSON objects with the symbols referencing the requested symbol.
Inputs: name_path (required), relative_path (required), exclude_kinds, include_kinds, +1 more. Returns: result
Retrieves information on all symbols/code entities (classes, methods, etc.) based on the given name_path,
which represents a pattern for the symbol's path within the symbol tree of a single file.
The returned symbol location can be used for edits or further queries.
Specify depth > 0 to retrieve children (e.g., methods of a class).
The matching behavior is determined by the structure of name_path, which can
either be a simple name (e.g. "method") or a name path like "class/method" (relative name path)
or "/class/method" (absolute name path). Note that the name path is not a path in the file system
but rather a path in the symbol tree within a single file. Thus, file or directory names should never
be included in the name_path. For restricting the search to a single file or directory,
the within_relative_path parameter should be used instead. The retrieved symbols' name_path attribute
will always be composed of symbol names, never file or directory names.
Key aspects of the name path matching behavior:
Trailing slashes in name_path play no role and are ignored.
The name of the retrieved symbols will match (either exactly or as a substring)
the last segment of name_path, while other segments will restrict the search to symbols that
have a desired sequence of ancestors.
If there is no starting or intermediate slash in name_path, there is no
restriction on the ancestor symbols. For example, passing method will match
against symbols with name paths like method, class/method, class/nested_class/method, etc.
If name_path contains a / but doesn't start with a /, the matching is restricted to symbols
with the same ancestors as the last segment of name_path. For example, passing class/method will match against
class/method as well as nested_class/class/method but not method.
If name_path starts with a /, it will be treated as an absolute name path pattern, meaning
that the first segment of it must match the first segment of the symbol's name path.
For example, passing /class will match only against top-level symbols like class but not against nested_class/class.
Passing /class/method will match against class/method but not nested_class/class/method or method. Returns a list of symbols (with locations) matching the name.
Inputs: name_path (required), depth, include_body, exclude_kinds, +4 more. Returns: result
Use this tool to get a high-level understanding of the code symbols in a file. This should be the first tool to call when you want to understand a new file, unless you already know what you are looking for. Returns a JSON object containing info about top-level symbols in the file.
Inputs: relative_path (required), max_answer_chars. Returns: result
Inserts the given body/content after the end of the definition of the given symbol (via the symbol's location). A typical use case is to insert a new class, function, method, field or variable assignment.
Inputs: body (required), name_path (required), relative_path (required). Returns: result
Inserts the given content before the beginning of the definition of the given symbol (via the symbol's location). A typical use case is to insert a new class, function, method, field or variable assignment; or a new import statement before the first symbol in the file.
Inputs: body (required), name_path (required), relative_path (required). Returns: result
Lists all non-gitignored files and directories in the given directory (optionally with recursion). Returns a JSON object with the names of directories and files within the given directory.
Inputs: recursive (required), relative_path (required), max_answer_chars. Returns: result
List available memories. Any memory can be read using the read_memory tool.
Inputs: no named parameters. Returns: result
Call this tool if onboarding was not performed yet. You will call this tool at most once per conversation. Returns instructions on how to create the onboarding information.
Inputs: no named parameters. Returns: result
Instructions for preparing for a new conversation. This tool should only be called on explicit user request.
Inputs: no named parameters. Returns: result
Reads the given file or a chunk of it. Generally, symbolic operations like find_symbol or find_referencing_symbols should be preferred if you know which symbols you are looking for. Returns the full text of the file at the given relative path.
Inputs: relative_path (required), end_line, start_line, max_answer_chars. Returns: result
Read the content of a memory file. This tool should only be used if the information is relevant to the current task. You can infer whether the information is relevant from the memory file name. You should not read the same memory file multiple times in the same conversation.
Inputs: memory_file_name (required), max_answer_chars. Returns: result
Replaces one or more occurrences of the given regular expression. This is the preferred way to replace content in a file whenever the symbol-level tools are not appropriate. Even large sections of code can be replaced by providing a concise regular expression of the form "beginning.*?end-of-text-to-be-replaced". Always try to use wildcards to avoid specifying the exact content of the code to be replaced, especially if it spans several lines.
IMPORTANT: REMEMBER TO USE WILDCARDS WHEN APPROPRIATE! I WILL BE VERY UNHAPPY IF YOU WRITE LONG REGEXES WITHOUT USING WILDCARDS INSTEAD!.
Inputs: repl (required), regex (required), relative_path (required), allow_multiple_occurrences. Returns: result
Replaces the body of the symbol with the given name_path.
Inputs: body (required), name_path (required), relative_path (required). Returns: result
Offers a flexible search for arbitrary patterns in the codebase, including the possibility to search in non-code files. Generally, symbolic operations like find_symbol or find_referencing_symbols should be preferred if you know which symbols you are looking for.
Pattern Matching Logic: For each match, the returned result will contain the full lines where the substring pattern is found, as well as optionally some lines before and after it. The pattern will be compiled with DOTALL, meaning that the dot will match all characters including newlines. This also means that it never makes sense to have .* at the beginning or end of the pattern, but it may make sense to have it in the middle for complex patterns. If a pattern matches multiple lines, all those lines will be part of the match. Be careful to not use greedy quantifiers unnecessarily, it is usually better to use non-greedy quantifiers like .*? to avoid matching too much content.
File Selection Logic:
The files in which the search is performed can be restricted very flexibly.
Using restrict_search_to_code_files is useful if you are only interested in code symbols (i.e., those
symbols that can be manipulated with symbolic tools like find_symbol).
You can also restrict the search to a specific file or directory,
and provide glob patterns to include or exclude certain files on top of that.
The globs are matched against relative file paths from the project root (not to the relative_path parameter that
is used to further restrict the search).
Smartly combining the various restrictions allows you to perform very targeted searches. Returns A mapping of file paths to lists of matched consecutive lines.
Inputs: substring_pattern (required), relative_path, max_answer_chars, paths_exclude_glob, +4 more. Returns: result
Activates the desired modes, like ["editing", "interactive"] or ["planning", "one-shot"].
Inputs: modes (required). Returns: result
Think about the collected information and whether it is sufficient and relevant. This tool should ALWAYS be called after you have completed a non-trivial sequence of searching steps like find_symbol, find_referencing_symbols, search_files_for_pattern, read_file, etc.
Inputs: no named parameters. Returns: result
Think about the task at hand and whether you are still on track. Especially important if the conversation has been going on for a while and there has been a lot of back and forth.
This tool should ALWAYS be called before you insert, replace, or delete code.
Inputs: no named parameters. Returns: result
Whenever you feel that you are done with what the user has asked for, it is important to call this tool.
Inputs: no named parameters. Returns: result
Write some information about this project that can be useful for future tasks to a memory in md format. The memory name should be meaningful.
Inputs: content (required), memory_name (required), max_answer_chars. Returns: result
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Scored across 25 tools
Most tools have clearly distinct purposes, such as find_file vs list_dir vs search_for_pattern. However, replace_regex and replace_symbol_body both modify file content, and the three think_about_* tools could be confused, though descriptions do differentiate them.
The vast majority use a consistent snake_case verb_noun pattern (e.g., find_file, read_file, write_memory). Minor deviations like 'onboarding' (noun) and 'check_onboarding_performed' (phrase) slightly break the pattern but remain readable.
With 25 tools, the server straddles the upper boundary of what is reasonable for its scope. The feature set is broad, covering file, symbol, memory, project, and meta operations, but the count feels heavy and could overwhelm an agent.
The server covers file read/create/edit/search, symbol analysis/editing, memory CRUD, and project onboarding, but lacks a dedicated file deletion tool (only delete_memory exists). Rename/move operations are also absent, which agents may need for full file lifecycle management.
Deterministic context layer for your codebase: change impact, blast radius, answers with receipts.
Code intelligence for LLMs. Analyze, search, and retrieve code from any public git repository.
Code intelligence for coding agents: semantic, AST, graph, and full-text search. 279+ languages.
The context platform for AI agents: code search, team decisions, and grounded project context.
