What it is
Serena, from Oraios AI, is an MCP server that gives a coding agent the code navigation and editing tools an IDE has. Instead of reading whole files and editing by line number, the agent can find a symbol, list who references it, replace a function body or rename a symbol across the project. Version 2.0.0 (13 October 2026) added a REPL-based interface in which the agent submits code rather than single tool calls, and 2.1.0 followed on 19 October; the repo has about 30,600 GitHub stars.
By default it runs on language servers (the engines editors use for go-to-definition), which the README says cover more than 40 languages. A paid Serena plugin for JetBrains IDEs is the second backend and adds move, inline, type hierarchy and an interactive debugger. It also has a simple memory system and optional file, search and shell tools, which are switched off by default inside harnesses such as Claude Code that already have them.
You install it with uv (uv tool install -p 3.13 serena-agent), run serena init, then point your client at it over stdio or HTTP. A web dashboard starts with each session on localhost:24282.
Install
uv tool install -p 3.13 serena-agentOur review
Most coding agents still explore a project with grep and full-file reads, and edit by matching text. Serena replaces that with symbol-aware operations: find a function, see every caller, swap its body, rename it everywhere. On a large codebase that means fewer tokens spent reading and fewer broken edits, which is the whole pitch, though the savings are claimed by the maintainers rather than independently measured.
The free language-server backend covers the core: find symbol, references, file outlines, diagnostics, symbol rename and symbol-level editing. The bigger refactors (move, inline, removing unused code) and debugging need the paid JetBrains plugin, so the free tier is narrower than the feature table first suggests, and some languages need extra language-server installs.
Security is the reason for a lower score. Serena’s own security page says it assumes the repository being worked on is trusted, and 2026 showed what that means: three published advisories, each ending in code execution. A critical template-injection flaw (fixed in 1.7.0) and a pickled symbol cache (fixed in 2.0.0) let a malicious repository run code merely by being activated, and an earlier bug let any web page reach the unauthenticated dashboard through DNS rebinding (fixed in 1.5.2). All are patched, but anyone pointing an agent at untrusted code should run Serena in a container, as the maintainers themselves advise.
Two defaults deserve attention: the dashboard starts with every session, and Serena reports anonymous usage data on startup unless SERENA_USAGE_REPORTING is set to false. The README’s "evaluation" quotes AI agents praising the tool, which is marketing rather than evidence, and it warns against installing from MCP or plugin marketplaces because their commands are out of date.
The licence matters for some teams: the application is GPL-3.0-or-later and contributions need a CLA. Running it as a local tool is fine; embedding it in a closed product is not. On small scripts the setup and extra tool descriptions cost more than they save, and an agent’s built-in grep and edit do the job.
Where it earns its place
A backend engineer on a large Java monorepo
Rename a widely used service method and update every caller
With the language-server backend the agent looks up references and renames the symbol in one call, instead of grepping and patching text across hundreds of files. It is free; the JetBrains plugin becomes necessary once the job grows into moving classes between packages.
A Python team lead rolling Claude Code out to 8 developers
Cut the tokens the agent spends reading whole files
File outlines and symbol lookup let the agent read one function instead of a 2,000-line module, and Serena switches off its own file and shell tools inside Claude Code to avoid duplicates. The cost is a uv install per developer and a dashboard window each session unless web_dashboard_open_on_launch is turned off.
An open-source maintainer reviewing outside pull requests with an agent
Navigate unfamiliar contributed code without exposing the workstation
Run Serena 2.0.0 or later inside a Docker container with only the checkout mounted, and keep contributor forks out of trusted_project_path_patterns. Its security model assumes trusted repositories, and two 2026 advisories showed a malicious repo could run code on activation.
Before you commit
- Setup
- Install uv, then uv tool install -p 3.13 serena-agent and serena init; client guides cover Claude Code, Codex, Claude Desktop and MCP-enabled IDEs, over stdio or HTTP. Some language servers must be installed by hand; others download on demand with pinned versions and SHA256 checks.
- Integrations
- Symbol tools for find, references, outline, declarations, implementations, diagnostics, rename and symbol-level edits, plus memories and optional regex search, file and shell tools. The JetBrains plugin adds move, inline, type hierarchy, dependency search and a debugger; Rider and CLion are unsupported.
- Data and privacy
- Runs locally with your user’s permissions and can edit files and run shell commands. Anonymous usage data goes to Oraios on startup unless SERENA_USAGE_REPORTING=false, and the dashboard can edit configuration and memories from localhost:24282.
- Cost at scale
- The language-server backend is free for any team size. The JetBrains plugin is a paid per-user licence with a 7-day trial; Oraios does not state the price in its own docs.
- Limits
- Language-server rename covers symbols only, not files or directories, and find-declaration generally fails for external dependencies. With the REPL interface every action appears as the same tool call, so client-side approval cannot tell a read from a write.
What's good
- Find, reference, rename and edit by symbol, not text
- Free language-server backend covers 40+ languages
- Works with Claude Code, Codex, Cursor, JetBrains assistants and others
- Actively developed; v2.1.0 released 19 October 2026
What's not
- Three code-execution advisories in 2026; assumes trusted repos
- Move, inline and debugging need the paid JetBrains plugin
- Dashboard and usage reporting on by default
- GPL licence and CLA matter for commercial embedding
Best for
Developers running coding agents on large, typed or multi-language codebases they trust, who want reliable cross-file navigation and renames.
Alternatives
Serena: common questions
- Is Serena worth it?
- Serena’s symbol-level tools are a real upgrade for agents in big codebases, but its trust-the-repository security model means untrusted code belongs in a container. We rate it 4.2 out of 5.
- How much does Serena cost?
- Free and open source (GPL-3.0-or-later; the SolidLSP part is MIT). The optional JetBrains plugin is paid, with a 7-day free trial; price is set on the JetBrains Marketplace.
- What are the downsides of Serena?
- Three code-execution advisories in 2026; assumes trusted repos; move, inline and debugging need the paid JetBrains plugin; dashboard and usage reporting on by default; gPL licence and CLA matter for commercial embedding.
- Who is Serena best for?
- Developers running coding agents on large, typed or multi-language codebases they trust, who want reliable cross-file navigation and renames.
- How much does Serena cost for a team or in production?
- The language-server backend is free for any team size. The JetBrains plugin is a paid per-user licence with a 7-day trial; Oraios does not state the price in its own docs.
- Is Serena safe for company data?
- Runs locally with your user’s permissions and can edit files and run shell commands. Anonymous usage data goes to Oraios on startup unless SERENA_USAGE_REPORTING=false, and the dashboard can edit configuration and memories from localhost:24282.
- What are the best Serena alternatives?
- The closest alternatives we have reviewed are Semble, code-review-graph, GitNexus and Context7.
Answers checked Oct 22, 2026.
Make Serena? Show this rating on your site
