Anvaya Harness · THE EXECUTION LAYER

The LightestAgentic Harness.

One Rust binary, no runtime dependencies, 25.6 MB median RSS and 17x lighter than Claude Code. That weight is the point: it is why hundreds of agents fit on one machine. Run it as a full-screen TUI or headless in a pipeline, on your own hardware through Ollama (no account or key) or against Anthropic and OpenAI-compatible gateways when a task earns it. Every turn is packed with experience from Anvaya Mind → and reports back what it learned, the same graph Claude Code and other agents can read over MCP.

25.6 MBmedian harness RSSPhase 1, 238 runs, 228 verified

~5 MBidle footprintCLI TUI, macOS measured

12 msfirst TUI framepty first-byte, 2026-08-14

256concurrent agentshard cap, local-first

ANV, HEADLESSanv 0.1.0

$anv init

✓ .anvaya/ initialized

$anv -m devstral:latest -p ollama

Keys 1-8 switch screens, Tab cycles, Ctrl+P palette, Y/N approves

Proof

Ran 25 Agents on a 2008 Laptop.

In 2008, Intel shipped the Core 2 Duo. It was a good chip then.

In 2026, we ran 25 concurrent Anvaya agent sessions on a laptop with that chip and 4 GB of RAM. Arch Linux. No crashes. No faults.

Each headful session cost ~21 MB. Headless, ~16 MB. Baseline, ~9 MB. The old instruction set fought us, so we migrated memory allocators to get around it. The network card bottlenecked under 25 concurrent inference calls, so we wrote a Rust router to multiplex them.

The point is not that Anvaya runs on old hardware. The point is that if it runs 25 agents on a Core 2 Duo, it runs hundreds on your laptop. And it is all local.

25concurrent sessions, no crashes, no faults

~21 MBper headful session, Core 2 Duo measured

~16 MBper headless session, Core 2 Duo measured

~9 MBbaseline footprint, Core 2 Duo measured

Rust routercustom multiplexer for 25 concurrent inference calls

Density

Hundreds of Agents. One Machine.

A 32-agent soak costs 103 MB total. 24 concurrent by default, 256 hard cap. 25 agents ran in parallel on a Core 2 Duo with 4 GB RAM. On a modern laptop, that is hundreds, all local, all on your disk.

103 MBtotal agent RSS, 32-agent soak measured

24 / 256default concurrent / hard cap

0crashes or faults in any soak or stress test

Localswarms run on your machine, not in the cloud

4 bundlesswarm, defect, mvp, refine, all local-first

By The Numbers

What’s Actually In It.

21Native tools, no plugin layer, no shelling out to Python

8TUI screens: Startup, Agent, Git, Monitor, MCP, Mind Viz, Terminal, Swarm

3Transports: Ollama, OpenAI-compatible (incl. Siemens preset), Anthropic

256Concurrent agents, hard cap, 24 by default

Eight Screens, One Keystroke Apart.

Everything the agent is doing stays visible. The terminal panel runs a split-tree workspace, so you can keep a build running beside the agent without leaving the TUI.

1 StartupModel, provider, and project resolution before the first turn.

2 AgentThe main loop, prompt, streamed reasoning, tool approvals.

3 GitStatus, diff, and file navigation without leaving the TUI.

4 MonitorLive view of the Mind daemon: workers, sockets, graph size.

5 MCPBrowse tools exposed by connected MCP servers.

6 Mind VizThe knowledge graph rendered, nodes, links, weights.

7 TerminalSplit-tree multi-pane shell. Ctrl+A toggles pane-command mode.

8 SwarmLive fleet board: agents, task state, per-agent RSS and tokens.

Twenty-One Tools, All Native.

Every tool is compiled into the binary. Destructive ones route through an approval gate unless you pass --yolo; --plan-only disables writes and execution entirely.

read_fileRead with range support

edit_fileTargeted string replacement

rewrite_fileFull-file replacement

patchApply a unified diff

globMatch paths by pattern

grepSearch file contents

search_in_fileScoped in-file search

list_dirDirectory listing

dir_treeRecursive tree view

create_file_or_folderCreate paths

delete_file_or_folderRemove paths

run_commandShell execution, gated by approval

run_scriptRun a saved script through the same sandbox

todowriteTrack multi-step work

webfetchFetch and read a URL

websearchSearch the web (keyless Exa/Parallel fallback)

lspPooled language-server queries

questionAsk the user before proceeding

taskDelegate to a sub-agent

mindQuery Anvaya Mind over Direct / IPC / MCP

monitorWatch long-running processes, up to 8 armed

Every Turn Is A Loop, Not A Prompt.

Between your input and the model sits a four-stage harness. The last two stages are what make experience compound, the loop closes back into Mind instead of ending at the reply.

01Intent EngineClassifies what the turn is actually trying to do before any context is assembled.

02Context PackerRequests experience from Mind under a token budget and packs it against the prompt.

03Turn LearnerRecords which injected nodes the turn actually touched, and the outcome, then sends it back.

04Session SynthesizerCondenses the finished session into durable nodes rather than a raw transcript.

It Runs Without You Watching.

Drop --no-tui and it becomes a normal Unix citizen: reads a prompt, writes to stdout, exits with a meaningful code. 0 on success, 1 on error, 2 when an approval was denied.

ANV, CI MODEstdin / stdout

$anv --no-tui --plan-only "what would change to add Redis caching?"

# read-only, never writes, never executes

$anv --no-tui --yolo "fix the auth bug in src/auth.rs"

$anv -s list

# resume any prior session by code, or --continue for the last one

Lightweight by measurement

Single-Digit MB At Idle. Milliseconds To Start.

Release build, macOS, footprint is truth (ps RSS includes mmap, excludes compressed pages, we print both). TUI idle ~5 MB footprint (12–15 MB ps RSS), headless setup ~3 MB, first frame in ~12 ms, idle CPU ~0% on 13 threads. A 17-hour live session held ~11 MB. Binary 14.3–18.7 MB depending on build, zero runtime dependencies. Measured 2026-08-14 protocol plus 2026-09-05 spot-check. The daemon side (warm ~50 MB, ingest bursts ~350 MB peak) lives in /benchmarks#footprint →

~5MBTUI idle footprint (ps RSS 12–15 MB)

~12msTo first TUI frame (--version 17 ms)

~0%Idle CPU, 13 threads

14–19MBStatic binary, opt-level 3 + LTO + strip

Run Agents That Fit On Your Laptop.

25.6 MB median RSS. 25 agents ran in parallel on a Core 2 Duo with 4 GB RAM. Hundreds on your machine. Zero cloud required on the Ollama path.

Get Started →Read the FAQ

Requires Rust/cargo to build from source. Linux and macOS today, Windows not yet supported. Pre-1.0, public beta. Pricing TBD.