JSPM

Mesh is a terminal-first AI engineering agent built to prove its work, not perform it.

Most agents stop when they sound done. Mesh keeps going until the change is verified — it runs your build and tests, scopes its own edits, and bundles the proof. Zero config: it ships with a shared LLM gateway, so there are no API keys to wire up.

npm install -g @trymesh/cli
mesh login        # opens a browser, hands the session back to your terminal
mesh              # start an interactive session

mesh login routes to Gemini 3.5 Flash through the shared Mesh gateway by default — no AWS/Google credentials needed. Requires Node 20+.

Why Mesh

• Proves its work. A convergence gate stops the model from spinning on read/search loops, an edit-scope guard blocks writes outside the task's directories (and never touches .git/, node_modules/, build output), and mesh pr turns a diff into a verifiable proof bundle.
• Zero-config auth. One mesh login, no keys. The shared gateway brokers every model call server-side.
• Token-efficient. Tool-spec compaction and transcript capsules keep prompts lean — typically far fewer tokens than a raw LLM loop for the same task.
• Codebase-aware. /index builds a workspace index; ask architecture, bug, or ownership questions across the repo with semantic + keyword search.
• Model-flexible. Swap between Gemini, Grok, GLM, Kimi, DeepSeek, Qwen, Llama and more with /model or --model <alias> — same session, same gateway.

Benchmarks

Head-to-head on a real multi-turn task (nextjs-full, 3 turns, Gemini 3.5 Flash) — Mesh vs a raw LLM loop on the same model and prompts. Both modes passed every turn; the difference is how many tokens reach the model and how long it takes.

nextjs-full · 3 turns · Gemini 3.5 Flash · identical pass outcomes

Input tokens sent to the model   (lower is better)
  Mesh   ██████████                            27.4K
  Raw    ████████████████████████████████████  94.6K     −71%

Wall-clock time
  Mesh   ██████████████████████                60.0s
  Raw    ████████████████████████████████████  96.7s     −38%

Per-turn input tokens — the gap widens as the session grows and a raw transcript balloons, while Mesh keeps the prompt lean:

Same answers, 71% fewer input tokens and 38% less wall-clock time. The savings come from tool-spec compaction and transcript capsules that trim the prompt instead of re-sending the whole history each turn. (Replay any time with mesh eval; raw figures use provider prompt-caching, so Mesh's true token savings are conservative here.)

Quickstart

mesh                      # interactive session
mesh "fix the failing auth test"   # one-shot: run a single task and exit
mesh init                 # first-run setup + repo briefing

Inside a session, type /help to list every command.

Commands

Useful in-session commands (full list via /help):

Models

Default is Gemini 3.5 Flash. Select any with --model <alias> or /model.

If the primary model fails, Mesh falls back automatically (Gemini 2.5 Pro → Kimi K2.6).

How verification works

Mesh treats "done" as something to demonstrate, not assert:

• Convergence gate — if the model loops on read/search tools without making progress, Mesh nudges it to converge, and ultimately forces a concrete answer or edit instead of endless exploration.
• Edit-scope guard — edits are confined to the paths your task implies; writes to .git/, node_modules/, .mesh/, and build directories are rejected outright.
• Proof bundles — mesh pr packages the diff with its verification result so a reviewer sees what changed and that it passed, not just a claim.
• Live watcher — mesh live --verify "<cmd>" runs your command on every change and flags regressions, no model in the loop.

Configuration

Most things work out of the box. Common knobs:

mesh --model pro "<task>"   # pick a model for this run
mesh --classic              # legacy scrollback UI
mesh --no-browser           # password login instead of browser flow

Auth & privacy

mesh login opens a browser to a Mesh-hosted page — the CLI never sees your password. On success, your session (a Supabase JWT) is stored in your OS keychain (with a ~/.config/mesh file fallback) and refreshed automatically. Every model request goes through the shared Mesh gateway, which injects provider credentials server-side; your prompts and code context are sent to the selected model provider to generate responses. Run mesh logout to clear stored credentials.

Requirements

• Node.js 20+
• A Mesh account (free tier available) — created during mesh login

Links

Website · Docs · Changelog

Contact

Questions, feedback, or partnership inquiries — reach the team:

• Edgar Baumann — edgar.baumann@try-mesh.com
• Philipp Horn — philipp.horn@try-mesh.com

License

See LICENSE.