Website · Docs

A terminal-first AI engineering agent. Built to prove its work, not perform it.

npm install -g @trymesh/cli
mesh login
mesh

mesh login opens a browser and hands the session back to your terminal. The shared Mesh gateway routes to Gemini 3.5 Flash by default. Requires Node 20+.

Headline behaviours

| | | | --- | --- | | Verifies | Every edit-turn runs pnpm test / pnpm typecheck / cargo check (auto-detected). No "tests pass" without an exit-0. | | Anti-thrash | After 3 consecutive edits without a passing verify, write tools are hard-blocked. The agent has to diagnose. | | Adaptive | Each prompt is classified — trivial gets a 900-byte prompt + 4 steps; complex gets the full scaffolding. | | Pre-fetch | Vague audits ("is X safe?") auto-read the top-3 relevant files before the LLM call. | | AST capsules | ts-morph for TS/JS, structured regex for Python/Rust/Go. Cached + invalidated on file change. | | OOM-safe | Pre-stat gate at 256 KB, slim-write fallback on huge indexes, 8 GB default heap. |

Models

Default is Gemini 3.5 Flash. Switch with /model inside the CLI or --model <alias> on the command line.

| Provider | Model | Alias | | --- | --- | --- | | Google | gemini-3.5-flash | flash, default | | Google | gemini-3.1-pro-preview | pro, gemini-pro | | Google | gemini-2.5-flash | 2.5-flash | | Google | gemini-2.5-pro | 2.5-pro | | Moonshot | kimi-k2.6 | kimi | | NVIDIA NIM | qwen / mistral / deepseek / llama variants | per-model alias | | Anthropic (via proxy) | claude-* | claude |

Commands

mesh                          # start interactive TUI
mesh --classic                # legacy scrollback REPL
mesh "<task>"                 # one-shot
mesh init                     # first-run setup + repo briefing
mesh doctor [fix]             # diagnose; --fix applies safe repairs
mesh /index                   # build the workspace index
mesh /tokens                  # context breakdown of the current session
mesh /undo                    # restore last file backup
mesh bench [--task id]        # run benchmark suite
mesh eval [--since ISO]       # replay historical turns; pass/fail scoring
mesh live                     # foreground watcher (verify + AI! triggers)
mesh pr [--intent …]          # build a PR proof bundle from current diff
mesh logout
mesh --version

Inside the interactive CLI, /help lists command groups.

Configuration

Environment variables. All optional — the defaults work.

| Variable | Default | Purpose | | --- | --- | --- | | MESH_FAST_PATH | off | Strict one-shot subprocess mode: minimal tool set, 4-step budget, 60 s wallclock. | | MESH_FORCE_TIER | auto | Pin scaffolding tier: trivial / small / complex. | | MESH_BENCH_RAW | off | Strip mesh's scaffolding for A/B comparison against a raw LLM. | | MESH_INDEX_TIMEOUT_MS | 8000 | Wallclock cap on ensureIndex. | | MESH_FIRST_INDEX_TIMEOUT_MS | 20000 | Wallclock cap on the first-run workspace index. | | MESH_TURN_TIMEOUT_MS | 60000 (fast-path only) | Hard kill on a single turn. | | MESH_DEBUG_LLM | off | Stream raw LLM request/response summaries to stderr. |

How a turn runs

prompt
  → classify complexity (trivial / small / complex)
  → maybe pre-fetch top-3 relevant files
  → build system prompt + capsule per tier
  → LLM call (streaming where supported)
    ↻ tool dispatch (read / grep / patch / run …)
  → verify command (extracted from package.json)
  → final answer with file:line citations

Workspace state

<project>/.mesh/
├── workspace.json        # meta marker; "indexed" | "partial" | "deferred"
├── index/                # ContextBudget + token-usage history
├── capsules/v1/          # AST capsule cache (mtime + content-hash keyed)
├── artifacts/            # tool-result artifacts (auto-pruned)
├── history/              # session transcripts
└── config.json           # workspace overrides

User-level state lives in ~/.config/mesh/. Both are size-capped (50 MB hard limit) and auto-recoverable.

Develop

pnpm --filter @trymesh/cli dev          # tsx watch
pnpm --filter @trymesh/cli build        # tsc + vite + minify
pnpm --filter @trymesh/cli typecheck
pnpm --filter @trymesh/cli test         # node:test, 276 tests
pnpm --filter @trymesh/cli smoke:release
pnpm --filter @trymesh/cli verify:release  # typecheck + build + test + smoke

License

Proprietary. See the root LICENSE.