Your AI agent opens file after file just to answer one question. CodeMesh hands it the exact lines instead. Same answer, a fraction of the tokens.
YOUR CODEBASE AS A GRAPH
Connects to the tools you already use
Measured on 98 real questions
98 real software-engineering questions on openclaw, a public repository anyone can clone and re-run. Same model, nothing excluded.
View Every Benchmark →Same repository, two very different bills
It searches, opens a dozen files, scrolls past most of them, then does the whole thing again on your next question.
One query walks the symbols, callers and imports, and returns the few spans that actually answer the question.
The same question, side by side
Token share is the published benchmark result: 25.3M tokens down to 1.23M across all 98 questions, so roughly 5% of the original spend remains. Per question the measured mean is 258,115 tokens down to 12,524.
In your terminal
Inside the editor
A public repository open in VS Code, Claude Code in the terminal. The question goes to the code-context tool, the graph walks from the repository to the lines that answer it, and the spans come back without a single file being opened.
Three real queries, three answers
Three questions against public repositories, answered the way your agent gets them: the tool that handled it, and the spans it returned with file and line. No crawl, and nothing opened that turned out to be irrelevant.
What comes back over MCPWhere your tokens actually go
Every session, your agent rediscovers the structure of a codebase that has not changed. CodeMesh already knows it, and simply tells the agent.
95.1% fewer retrieval tokens.
And with less irrelevant code in the context window, the model has more room left to reason about your actual problem.
Estimate your own saving
Estimates are illustrative and show potential savings on eligible context/retrieval usage only, not your entire AI bill. Actual savings depend on model pricing, coding-agent behavior, repository size, task type and the proportion of usage attributable to code retrieval and context loading.
Fast, and never stale
Cutting tokens is no use if your agent then sits waiting. Structural graph queries return in roughly a millisecond, so the context arrives faster than reading the file would have.
Lower cost. Less context. Fast retrieval.
Save a file and only that file is reparsed, not the whole index. The graph keeps pace with your repository, so your agent never answers from last week's code.
Backed by blake3 manifests and incremental tree-sitter parsing
Install to first answer
Install the CodeMesh extension in Cursor, Devin or VS Code.
CodeMesh parses it and keeps the graph in step with every save.
Point Claude Code, Cursor or any MCP client at CodeMesh. One config block.
No editor needed
For Claude Code, Codex, Opencode and anything else that runs in a shell. One install, then sign in, sync a repository and point your agents at it.
Read the CLI guideWhat your agent does instead
The panel, step by step
The extension panel, walked through the way a fresh install goes. The tags say what each part does.
The name and address on the account, and the plan it is on. One click from any screen, on the avatar in the header.
Seats taken, storage against your allowance, and queries so far today. The same figures the console shows, without leaving the editor.
Manage plan opens billing in the browser. Nothing about your subscription is edited inside the editor.
It pauses syncing and agent queries on this machine and removes the tool guard. Your graph and the agent's registration stay.
The name and address on the account, and the plan it is on. One click from any screen, on the avatar in the header.
The real extension, running, on a loop. Not clickable here: install it to use it.
For engineering teams
Nothing to switch and nothing to relearn. CodeMesh sits underneath the tools your team already has and makes the context they send much smaller.
One developer wasting tokens is a rounding error. Fifty is a line item on your bill. It happens every session, on every repository, and so does the saving.
More prompts. More features. Same AI budget.
Illustrative example, not a savings guarantee
Built for real repositories
Ask about a function and get that function back, with its file and line numbers.
Callers, imports and call edges come back from one query instead of ten tool calls and a lot of guessing.
The graph persists, so a brand new chat does not begin by relearning your project from scratch.
A smaller context window leaves the model more capacity for the actual task.
Incremental sync reparses what changed, never the whole repository.
Claude Code, Cursor, Devin, VS Code and any MCP client. No proxy, no wrapper, no new editor to learn.
After you sign in
The five screens you work in, running right here. Pick one to see what it tells you.
Where you land. Whether every repository synced clean, how large the graph has grown, how many seats are in use, and which agents currently hold read access.
Example workspace (Northgate Systems). The reduction shown is the published benchmark's 95.1%.
How the graph is built
A continuously updated structural graph of your repository. That is the whole trick, and it is why retrieval costs so little.
Pricing
The heavier your agent use, the more of your bill is simply repeated reading. That is the part CodeMesh takes out.
If CodeMesh saves more in unnecessary AI context usage than the subscription costs, it pays for itself. Calculate My Savings →
Security
Access is scoped by the credential, never by anything the client sends. So cutting your AI costs never means giving up control over your source code.
Connect a repository, point your agent at it, and see what the same work costs afterwards.
Cut token consumption by 95.1%.
Code more. Spend less on AI.