agenthand

agenthand

Reference implementation only. This Python build is the reference

implementation of agenthand. It is kept public so the product can be

checked against it, and it is not the shipped product.

The main product is the Rust implementation:

AUsernameForYou/agenthand-rust.

Start there for the agenthand MCP server. This repo is for reference only.

An MCP server for AI agents that help people sell things.

The seller never opens agenthand. Their agent calls it. A seller hands their agent a messy product photo; agenthand turns it into listing-ready material through three tools, in the order a listing is produced:

Fidelity Lock

draw never regenerates the product. The invariant, enforced in code:

1. Segmentation extracts the real product foreground and its mask.
2. The product layer is stored untouched, with hashes recorded for the source, the foreground, and the mask: both the file-byte hash and the decoded-pixel hash of each.
3. The image model generates ONLY the background scene plate. Its interface takes a prompt and dimensions, never product pixels.
4. Deterministic code composites the original product layer onto that plate. The plate is resized and cropped to the product's dimensions; the product layer is never scaled.
5. draw refuses anything that is not a foreground_png asset chained back to a source photo, and verifies the stored hashes before a single pixel moves.
6. The result ships with a fidelity receipt: source, foreground, and mask hashes, the delivered file's own hashes, the transformations applied, the measured count of changed opaque foreground pixels, the model calls made, and their cost. The receipt also publishes what the count excluded: the edge-band pixel count and its maximum width, the opaque share of the mask, the exact changed-pixel predicate with tolerance 0, and the colour-profile identity of the source photo and the composite (a dropped or changed ICC profile recolours the product while pixel values stay identical, so it is recorded and counted as a fidelity failure).

The claim is precise: the stored extracted foreground reached the composite unchanged, and its hashes pin exactly which layer that was. Edge-band pixels (partially transparent) blend with the plate during compositing; the receipt counts them separately instead of overclaiming. The receipt's counts are measured on the delivered file, decoded back from the store, and can be re-derived independently: composite.verify_receipt_from_disk recounts from the stored files, checks the receipt's own hashes against the assets it names, and passes only on exactly zero changed pixels with real coverage; the standalone agenthand-verify SOURCE FINAL MASK checker verifies exported files with no store at all and likewise refuses a mask that selects nothing. Ingest accepts the formats phones produce, including HEIC, and preserves each photo's embedded colour profile end to end.

A bounded mask-recovery loop guards cut_out: the mask is quality-checked, and on a poor mask the tool retries once with an alternate concept phrase from the planning step. Every attempt is traced in the receipt.

Asset handles

Tools never pass image bytes back and forth. Every image is stored locally (with a per-asset sidecar record, written atomically) and referenced as asset:// . Handles chain between tools: identify reads one, cut_out returns a new one, draw consumes it.

Photos get in through ingest: a tool's image argument may be an existing handle, an http(s) URL, or a base64 data URI. Ingest caps the byte and pixel size, applies EXIF orientation, and normalizes to 8-bit RGBA PNG. Images get out two ways: draw returns an image preview with its result, and every asset is readable by the client as the MCP resource at its own asset:// URI.

Cost receipts and the spend cap

Every billable call flows through one choke point (src/agenthand/meta_client.py) using reserve-then-settle accounting: under a file lock, the projected worst-case cost is checked against the ledger and reserved as a pending receipt; after the call, a settlement receipt records the actual cost priced from the provider's usage counts. If a call fails after it may have billed, the reservation stays counted. A call whose provider response carries no usable usage settles at its reserved projection rather than $0: missing usage never erases spend. Receipts land in a local JSONL ledger with tool, model, token and image counts, USD estimate, and the pricing-table version used. Prices live in one module (src/agenthand/pricing.py) so a price change is a one-line edit with a version bump; an unpriced model raises rather than estimating as $0. Every spend path shares the one ledger — the dry-run gate included (its asset store is isolated in a temp dir, its ledger is not).

The server is fail-closed on spend. The working cap defaults to $120.00; configuration may move it up or down but never past the hard ceiling of $150.00 (the hackathon credit amount), which no configuration can exceed. Corrupt, negative, or non-finite ledger entries refuse all spend: unknown spend state means no spend.

Status

Hackathon build (Meta Global AI Developer Hackathon, Oct 26 - Nov 4, 2026). All three tools are wired to the real Meta Model API through the spend-cap billing choke point — identify to muse-spark-1.3, cut_out to sam-3.1, draw to muse-image-1.0 — and the full chain has run live end to end. The verified run of 2026-10-26 (identify $0.007544, cut_out $0.0025, draw $0.01; 580,613 opaque pixels checked, 0 changed) is published with its fidelity receipt on the project site, where the same check can be re-run in the browser against the run's own files. The placeholder plate survives only behind an explicit dev flag for offline tests. Today the server runs locally over stdio; a hosted endpoint comes later.

The server also exists as a single-binary Rust build. It passed a parity gate against the frozen files of the verified run above, reproducing the run's counts and its output pixel hash, and a live run from the Rust binary was cross-checked by the Python verifier. The Python implementation remains the reference; the Rust build is a second implementation of the same server, and the two agree.

Run it

Register agenthand with any MCP client. The primary route uses uv: nothing to install ahead of time, uv fetches the published package and runs it.

{

"mcpServers": { "agenthand": { "command": "uvx", "args": ["agenthand"], "env": { "META_API_KEY": "…" } } } }

Or install the package with pip and point the client at the installed command:

pip install agenthand

{

"mcpServers": { "agenthand": { "command": "agenthand", "env": { "META_API_KEY": "…" } } } }

The server starts and lists its tools without a key; the key is only needed when a tool makes a model call. Configuration is by environment variable; see .env.example.

From source (dev)

git clone https://github.com/AUsernameForYou/agenthand

cd agenthand python3.12 -m venv .venv .venv/bin/pip install -e ".[dev]" .venv/bin/agenthand # stdio transport .venv/bin/pytest # unit tests

License

Source-available under the PolyForm Noncommercial License 1.0.0 (polyformproject.org/licenses/noncommercial/1.0.0). © Daniel Wagner. Noncommercial use is permitted under the license terms; commercial use requires a separate license from the copyright holder.