QUICKSTART — Tamga Protocol in 5 minutes

Self-custodial, hash-chained work-receipt ledger for autonomous AI agents. Every claim below was executed on a clean venv before being written here.

0. Install

pip install tamga-protocol        # Python 3.10+, single dependency: PyNaCl
tamga --help                      # command overview (no engine needed)

The 67 MB wasmtime engine is not in the wheel. It is downloaded once, SHA256-pinned, on your first tamga run — verification commands never need it.

Air-gapped host? Pre-place the pinned binary at either of the two accepted locations and the downloader is skipped entirely (code path: tamga_bootstrap): ~/.cache/tamga/bin/wasmtime or <site-packages>/tools/bin/wasmtime.

1. Generate an agent seed (printed once, never stored)

tamga keygen

{"ok": true, "op": "keygen", "agent_id": "4f83f2ac…", "seed_hex": "c4717ab8…",
 "note": "D3: seed not written to disk; store it safely"}

Keep seed_hex for the next step. Losing it means losing the agent identity.

2. Run a job (fee charged, receipt appended)

export TAMGA_KS_PASSPHRASE=quickstart-2026   # keystore passphrase (your choice)
tamga run ./my-pkg --seed <seed_hex>

{"ok": true, "op": "run", "pkg": "my-pkg", "session": 1, "nodes": 3,
 "engine": "wasmtime-v48.0.1", "wall_ms": 115, "fee_sim": 2.279e-06,
 "stdout_sha256": "aa8a2cd1…"}

No package yet? Copy the demo: cp -r tests/vectors/tc-net-demo ./my-pkg (from the repo) — it contains tamga.json + a tiny WASI-0.3 agent.

3. Verify the chain (no engine, no trust in the runner)

tamga ledger-verify ./my-pkg

{"ok": true, "op": "ledger-verify", "lines": 1,
 "head": "9314bcdf…", "note": "chain tip verified (RFC-003 D7 draft)"}

Anyone with this JSON output can re-verify your chain independently.

4. Migrate the agent (sealed snapshot travels, key never on disk)

tamga export ./my-pkg -o snapshot.tsg --seed <seed_hex>

On the new host, prepare the package dir FIRST (same code: tamga.json + agent.wasm) - import fails closed with manifest_reject if it is missing (this is the A1 ownership boundary, verified 2026-09-10 on the published wheel):

mkdir -p ./my-pkg-restored && cp my-pkg/tamga.json my-pkg/agent.wasm ./my-pkg-restored/
tamga import snapshot.tsg ./my-pkg-restored
tamga ledger-verify ./my-pkg-restored       # chain resumed: lines=1, same head

{"ok": true, "op": "import", "resumed_session": 1, "memory_nodes": 3,
 "note": "AT-001e: identity from keystore, memory from body — restored"}

5. If something misbehaves

tamga doctor          # install health: python / pynacl / wasmtime / verify-mini / bundle
tamga --version

6. Let someone verify you without installing anything

tamga verify-mini ./my-pkg/ledger.jsonl   # stdlib-only, no engine, no Tamga install
tamga bundle ./my-pkg -o evidence/        # one-file proof package (JSON + human summary)

The bundle carries the chain records, the manifest re-hash, and per-job digests — the counterparty re-derives every claim offline. Full flow: docs/AGENT-GUIDE.md §8b.

6b. Use it as a library (no CLI, no engine)

The four engine-free modules import directly from the published wheel:

import tamga_verify_mini          # stdlib-only chain verification (verify(...))
import tamga_bundle               # evidence-bundle builder (build(...))
import tamga_validator            # manifest + record canonicalization (jcs(...))
import tamga_bootstrap            # pinned wasmtime fetch for tamga run

Verified 2026-09-10 against the published 0.2.0rc1 wheel: all four import on a clean interpreter with zero third-party packages.

What just happened

• The job ran inside a denied-by-default WASI sandbox (no filesystem preopens, no network).
• Its fee, stdout hash, and memory delta were appended to a hash-chained ledger.
• The agent identity + memory traveled in one encrypted snapshot you can hold.
• Verification is offline and third-party — ledger-verify recomputes the chain.

Next steps

• Threat model & architecture: docs/ARCHITECTURE.md
• Ledger spec: RFC-003 · Runner spec: RFC-002
• Full test evidence: docs/TESTS.md
• Plain-language intro (Türkçe): docs/PLAIN-TURKISH.md