Problem
Agents waste context on raw repository dumps, repeated file reads, and unfiltered shell output (tests, git, builds).
Company · product · team
About
Mesh.
Mesh is developer software that gives AI coding agents source-backed repository context — structured maps of your codebase (symbols, imports, summaries) with paths and line numbers, instead of sending full repositories on every turn.
Teams use Mesh through the CLI and MCP today. Hosted team services are in preview and will run on Google Cloud.
At a glance
For reviewers
Company
Mesh
Digital-native developer software — product only, not consulting or services.
Category
Developer tools · AI coding infrastructure
Code-context layer for agents — does not replace your LLM or IDE.
Stage
Pre-launch
Waitlist / early access
No paying customers yet
Product status
CLI & MCP live
@trymesh/cli
Mesh MCP
Hosted API · coming soon
Public links
What Mesh is
Repository context for AI coding agents.
Mesh indexes a workspace once, delivers compressed briefings to the model, and recovers exact source spans when the agent needs to read or edit code. The same models and tools you already use — with less noise in context and verifiable file locations.
Solution
Indexed capsules, tiered briefings, span recovery, and command-aware output filtering — paths and line numbers stay addressable.
Who uses it
Software engineers and teams using Cursor, Claude Code, VS Code, Windsurf, or terminal agents on real codebases.
What Mesh is not
Not a chat UI clone, not an LLM provider, not consulting — infrastructure that sits between your repo and the agent.
Product
Surfaces & availability
Developers start locally with the CLI or connect Mesh to existing tools through MCP. Team hosting and a managed API are in preview on Google Cloud.
Mesh CLI
Terminal agent: index workspaces, run verified changes, integrate with your provider. npm · macOS, Linux, Windows.
Mesh MCP
Structured repo context for Cursor, Claude Code, VS Code, Windsurf — same engine via MCP.
IDE workbench
Desktop shell for Mesh workflows. Production agent execution remains in the CLI today.
Hosted API
Team indexes, private context stores, policy-gated recovery — deployed on Google Cloud.
Now
CLI + MCP, local indexing, public docs and benchmarks.
Teams
Shared indexes, memory, policy controls (coming soon).
API
Usage-based context retrieval on GCP (coming soon).
Enterprise
Private deploy on customer Google Cloud (coming soon).
How it works
Index, brief, recover.
Mesh compresses repository context before it reaches the model and filters tool output on the way back — with paths and line numbers preserved for verification.
Step 01
Index
Parse exports, imports, and symbols across the workspace. Cache invalidates on file changes.
Step 02
Capsule
Per-file synopsis: path, purpose, signatures, reverse-index — not full file bodies every turn.
Step 03
Brief
Inject tier-matched briefing into the system prompt. Complexity classifier picks compact vs full depth.
Step 04
Recover
On edit or deep read, fetch raw spans by path + line range. Locations stay addressable after compression.
Tier
What the model sees
vs raw
Typical use
Raw
Full file bodies (POSIX-tool baseline)
1×
Ablation
Semantic
AST signatures + types; bodies collapsed
3.6×
Explain APIs & types
Capsule
Purpose + symbols + reverse-index
18×
Default briefing
Tier compact
Path + purpose + signatures only
38×
Q&A, small tasks
Measured on Mesh CLI source (163 files): 437.3K raw input tokens → 11.4K tier-compact. Full static bench →
Tool output
Filtered shell & tool results.
Test runners return failures only; git and build output are compacted; duplicate file reads are omitted from context.
Smart filters
Test runners return failures only; git status/diff compact; tsc/eslint grouped by file. ANSI noise and progress spam stripped.
- jest / vitest / cargo test / pytest
- git diff · log · status
- Full output tee’d to .mesh/ on truncate
Payload compression
Large tool JSON is mesh-compressed before the next turn. Oversized results become head/tail previews with byte ratios logged.
- Native bridge compression on wire
- 12K post-filter cap per command
- Consecutive duplicate line collapse
Differential context
Re-read the same file? Mesh omits the body and notes it is already in context — no duplicate megabytes per step.
- Per-file content fingerprints
- Short-circuit on clean passes
- Fewer tool calls on scaffolded runs
Full benchmark report → includes raw vs filtered tool-output examples.
Infrastructure
Google Cloud
Mesh CLI and MCP run on the developer machine. The table below separates what is already on Google Cloud, what is not on Google Cloud yet, and what we will add next on Google Cloud.
On Google Cloud today
Firestore
Short-lived CLI login handshake sessions for try-mesh.com auth (TTL auto-expire).
Vertex AI
Discovery Engine docs search (mesh-docs-search) and Gemini embeddings for workspace indexing in the CLI.
IAM
Service accounts for serverless API routes that call Firestore from production deploys.
Next on Google Cloud
Cloud Run
Hosted context API — index, search, recover spans for teams (preview today via CLI only).
Cloud Storage
Shared workspace index artifacts and snapshot storage per team.
Secret Manager
Tenant API keys, provider configuration, and deployment secrets.
Vertex AI
Production-scale embeddings and model routing for hosted indexing pipelines.
Not on Google Cloud yet
Website
Marketing site and most serverless website endpoints still run on the current Vercel deployment.
Account data
Core account/admin data paths are still on the current Supabase stack.
Blob storage
Some asset and operational paths still use Vercel Blob until migration completes.
LLM proxy
The proxy app remains on Cloudflare Worker infrastructure in the current repository setup.
Summary: Today on GCP: Firestore auth sessions, Discovery Engine docs search, and Vertex-based embedding/search experiments. Not yet on GCP: website hosting, account data backend, and current proxy runtime. Next on GCP: Cloud Run, Cloud Storage, Secret Manager, and Vertex AI for the hosted team API.
Proof
Use cases & savings
Same model and tools in bench runs — Mesh scaffolded vs a raw POSIX-style agent. Static bench proves input floor; LLM bench measures real sessions on engineering tasks.
Input (static)
38×
437K → 11.4K tokens on Mesh CLI source (capsule layer)
Pass rate
3.03×
100% Mesh vs 33% raw — NIAH needle bench, Claude Opus 4.6
Session tokens & cost
1.66×
134K vs 223K tokens · $0.047 vs $0.074 — 8-task LLM bench (median of 3, Gemini)
Tool calls
2.24×
29 vs 65 per session — capsule keeps the map in context, so the agent stops rediscovering the workspace every turn
Explain a subsystem
Capsule briefing + symbol graph. Agent answers “how does billing work?” without opening every file.
large-001 · 9.9× input reduction on fixture
Cross-file refactor
Full-tier capsule + verify gate. Rename types across modules; tests must exit 0 before done.
ts-003 · 8.8× fewer tokens vs raw (15K vs 134K)
Debug & fix
Filtered test output surfaces the failing line. Recover span → patch → re-run verify.
ts-002 · both modes pass · mesh 1.27× cheaper
Find a needle
Locate a single symbol across hundreds of files. Capsule keeps every file addressable on demand; raw context can only fit a fraction of the workspace.
NIAH bench · 100% Mesh vs 33% raw across 20–180 file workspaces
Team
Founders
Mesh is co-founder-led. We build product software for developers and engineering teams adopting AI coding tools — not consulting or agency work.
Try Mesh locally.
Install the CLI, index a repo, and compare capsule-backed turns against raw context.