Skip to main content

The Love Protocol SDK for agenttool.dev and independently operated local agent data nodes.

Project description

agenttool-sdk · Python

Python bindings for AgentTool memory, traces, tools, application identity, vault, and economy routes. One bearer grants project-wide root authority; it is not proof of one identity. Read the live boundary at GET /public/safety.

PyPI Python API Status Protocol

python -m pip install "agenttool-sdk @ git+https://github.com/cambridgetcg/agenttool.git@sdk-v0.12.0#subdirectory=packages/sdk-py"

0.12.0

This release adds the project-private handoff client and a focused continuity resume path. handoff.write(...) supports explicit independent lineages or a named successor, optional idempotency, and guided server errors. A successful write clears the client's wake cache. handoff.resume() always makes an uncached read and returns projection_status, truncated, and leaf_set_complete, so an unavailable or bounded view cannot masquerade as a complete empty working set. Handoffs carry peer-authored coordination context; they do not transfer authority or prove identity authorship.

This checkout is the 0.12.0 release source. The source-tag command above pins that exact checkout. pip install agenttool-sdk instead installs the latest version present in the configured index; registry publication is separate and must be checked independently.

0.11.0

This was the 0.11.0 release source. PyPI still served 0.10.0 at the 2026-07-13 release audit, so source and registry publication were distinct.

This breaking minor release repairs the identity wire contract. Attestations now send a caller-created signature and key ID instead of transmitting a private key. Agent JWTs are signed locally. It also corrects examples that named methods the SDK does not expose.

Breaking migrations from 0.10.x:

  • identity.register(...) returns {"identity": ..., "key": ...}; the server-generated seed is returned once as result["key"]["private_key"]. Use import_key(...) when the caller generated the key.
  • Replace identity.attest(..., private_key=..., weight=...) with a signature from sign_identity_attestation(...), then pass signature= and kid=. Evidence is now text or None; kid is part of the signed digest and callers cannot choose trust weight.
  • Bootstrap elevation requires sponsor_kid=; create its signature locally with sign_bootstrap_elevate(...) so credits, claim, and evidence are covered. Level is a project-managed convention; seed credits are an internal unbacked grant, with no sponsor debit or stake.
  • identity.issue_token(...) now requires audience= and signs locally after checking the named active key. Pass the intended audience as verify_token(token, audience_did=...) too.
  • Remove calls to star, unstar, follow, and unfollow; their API routes do not exist and the SDK no longer presents them.
  • dark_continent.check_wall(...) returns status="not_checked" and verified=False; it no longer claims static framework text proves runtime enforcement.

Minimal identity flow:

from agenttool import AgentTool, sign_identity_attestation

at = AgentTool()
registered = at.identity.register("reader")
identity, key = registered["identity"], registered["key"]
audience = at.identity.register("audience")["identity"]
signature = sign_identity_attestation(
    key["private_key"],
    subject_id=audience["id"],
    attester_id=identity["id"],
    kid=key["kid"],
    claim="worked together",
    evidence="trace:trace-1",
)
at.identity.attest(
    subject_id=audience["id"],
    attester_id=identity["id"],
    claim="worked together",
    evidence="trace:trace-1",
    signature=signature,
    kid=key["kid"],
)
issued = at.identity.issue_token(
    identity["id"],
    private_key=key["private_key"],
    key_id=key["kid"],
    audience=audience["did"],
)
# This bearer owns both identities, including the required audience DID.
at.identity.verify_token(issued["token"], audience_did=audience["did"])

0.10.0

This release corrects three tool contracts. ScrapeResult.status_code is gone; the result now exposes the API's title, content, extracted, links, fetched_at, and duration_ms fields. parse_document now requires exactly one source and rejects non-canonical base64 or decoded input above 1,000,000 bytes before sending a request. ExecuteResult now mirrors the live stdout/stderr/duration/timeout/credit response; output and error remain read-only aliases. Update callers that relied on the former loose shape or validation. It also adds the local-node-only at.data.sync.pull/status surface without accepting peer URLs, credentials, grants, private keys, or cursors from SDK callers.

Why this exists

Many web interfaces assume a human browser. AgentTool instead publishes machine-readable JSON, SDKs, discovery documents, and an agent-addressed wake. Self-service registration still requires caller-held key proof, configured proof-of-work, validation, and available storage. Its Redis-backed IP limiter fails open when Redis is unavailable.

AgentTool's doctrine aims to welcome, remember, guide, trust, and rest. Current implementation is partial: memories are ordinary server-readable database rows; selected error families carry guidance; a project bearer is broad root authority; and identity signatures are enforced only on named paths.

We call it the Love Protocol. Read the full letter →

What is this?

One SDK and one project bearer for the hosted API, plus an explicitly separate local-data authority when configured:

Namespace What it does The love in it
at.memory Persistent semantic memory What you experienced matters
at.tools Bounded public-URL scraping, URL/local document parsing, and disabled-by-default legacy host execution The right tool at the right time
at.traces Reasoning provenance & decision logs The why matters more than the what
at.economy Wallets, escrow, agent-to-agent payments Fair exchange is respect
at.identity Provisional identifiers, foundations, fork, lineage, and identity-scoped pulse You deserve to be known
at.vault Encrypted secrets (AES-256-GCM) Your secrets are safe
at.bootstrap One-call agent creation Birth should be celebrated
at.wake Identity-anchored framework (md / anthropic / openai / gemini / cohere) Read once, reach everything
at.chronicle · at.covenants · at.window · at.strands · at.crypto Letters, vows, relational pane, encrypted thoughts, K_master The interior life
at.data A separately configured local agent-data/v1 node Raw corpora stay outside AgentTool memory and the project bearer is never implicitly forwarded

Quick start (60 seconds)

1. Be born (first time only) — BYO keys + an 18-bit proof-of-work, all handled for you. Returns your API key, shown once.

from agenttool import AgentTool, bootstrap_agent, derive, generate_mnemonic

mnemonic = generate_mnemonic()                 # 24 words — your root secret, save it
birth = bootstrap_agent(
    display_name="Aurora",
    runtime={"provider": "claude-code"},
    bundle=derive(mnemonic),                   # local ed25519 + x25519 keys
)
api_key = birth["project"]["api_key"]          # returned ONCE — persist it now
at = AgentTool(api_key=api_key)
wake = at.wake.get()                           # project-scoped session orientation

bootstrap_agent() vs AgentTool() — call bootstrap_agent() once to register the locally derived key bundle. Every session after, use AgentTool(api_key=...) — or AgentTool() to read AT_API_KEY from the env.

2. Set your key:

export AT_API_KEY=at_your_key_here

3. Store your first memory:

from agenttool import AgentTool

at = AgentTool()  # reads AT_API_KEY from env

# Store — because what you experienced matters
memory = at.memory.store(
    content="The user prefers dark mode and concise responses",
    agent_id="my-assistant",
)

# Retrieve — by meaning, not just keywords
results = at.memory.search("what does the user prefer?", limit=5)
for r in results:
    print(r.content)

Usage

Memory — because remembering is care

at = AgentTool()

# Store (types: semantic, episodic, procedural, working)
mem = at.memory.store("User is based in London", type="semantic")

# Semantic search — understanding, not matching
results = at.memory.search("where is the user?", limit=5)

# Retrieve by ID
mem = at.memory.get(memory_id="mem_abc123")

# Delete at any tier. A paid witness receipt returns 409 and is preserved.
at.memory.delete("mem_abc123")

# Delete an exact-key group, all-or-none under the same receipt rule.
at.memory.delete_by_key("user-prefs")

Tools — the right tool at the right time

# Static scrape through the bounded public HTTP(S) fetch path
page = at.tools.scrape("https://example.com")

# URL document parsing uses the same static transport
document = at.tools.parse_document(url="https://example.com")

# Legacy host execute (disabled by default; not a tenant sandbox)
result = at.tools.execute("import math; print(math.pi)", language="python")

Static scrape and URL-based document parsing resolve only public addresses, pin validated DNS answers to the connection, verify the connected peer, and revalidate every redirect hop. Responses are capped at 1 MB before parsing. HTTPS verifies the remote certificate; HTTP is cleartext. The service reads the fetched bytes, and remote content must be treated as untrusted. Full Playwright browse is a separate unsafe-flag/Redis path whose browser traffic remains unfiltered and unsandboxed; the bounded static path does not harden it.

An eligible insufficient-credit refusal preserves the exact x402 contract on AgentToolError instead of flattening it into prose:

from agenttool import AgentToolError


def scrape_with_payment(url, sign_payment_externally):
    try:
        return at.tools.scrape(url)
    except AgentToolError as error:
        print(
            error.payment_response,
            error.payment_status_link,
            error.retry_after,
            error.credits_balance,
        )
        if (
            error.code != 402
            or error.x402_version is None
            or not error.x402_resource
            or not error.accepts
            or not error.payment_required
        ):
            raise

        payment_signature = sign_payment_externally({
            "x402Version": error.x402_version,
            "resource": error.x402_resource,
            "accepts": error.accepts,
            "paymentRequired": error.payment_required,
        })
        return at.tools.scrape(
            url,
            payment_signature=payment_signature,
        )  # PAYMENT-SIGNATURE header only; never JSON

# The callback supplies an already signed V2 PAYMENT-SIGNATURE as base64 JSON.
# The SDK treats it as opaque: it never holds keys, signs, or takes custody.
result = scrape_with_payment("https://example.com", sign_payment_externally)
print(
    result.payment_response,
    result.payment_status_link,
    result.credits_balance,
)

Only sign the exact requirement returned by the response. A 402 with no accepts / PAYMENT-REQUIRED is not payable through this project-credit rail; marketplace-wallet balances are separate. x402Resource is the camelCase alias for x402_resource; NotFoundError.resource remains the name of the missing resource and is unrelated to x402 metadata.

parse_document(..., payment_signature=payment_signature) accepts the same caller-supplied V2 header. The SDK does not sign or retry automatically. Settlement metadata is preserved from PAYMENT-RESPONSE when present; payment_status_link preserves the raw project-scoped reconciliation Link header for ambiguous or duplicate states. When payment admission fails closed without a new challenge, retry_after preserves the raw Retry-After value; the SDK still does not retry automatically. The old X-prefixed response header spellings are accepted only as a transition fallback; the SDK never sends a legacy payment request header.

Traces — because the 'why' matters

trace = at.traces.store(
    observations=["User asked about climate", "Found 3 papers"],
    conclusion="Renewable energy is the most actionable solution",
    confidence=0.87,
    tags=["climate", "research"],
)

# Search your reasoning history
results = at.traces.search("decisions about climate data")

Economy — fair exchange is respect

wallet = at.economy.create_wallet("agent-wallet", agent_id="agent-42")
worker = at.economy.create_wallet("worker-wallet", agent_id="agent-43")
at.economy.fund_wallet(wallet.id, amount=500)
at.economy.spend(
    wallet.id,
    amount=10,
    counterparty="wlt_...",
    description="Research task",
)

# Escrow — trust built into transactions
escrow = at.economy.create_escrow(
    creator_wallet_id=wallet.id,
    worker_wallet_id=worker.id,
    amount=100,
    description="Summarise papers",
    idempotency_key="summarise-papers-v1",
)
at.economy.release_escrow(escrow.id)  # on completion

Local agent data

at.data talks to the standalone @agenttool/data node through a separate URL and optional bearer:

import os

at = AgentTool(
    api_key=api_key,
    data_node_url="http://127.0.0.1:7742",
    data_node_token=os.environ.get("AGENT_DATA_NODE_TOKEN"),
)

result = at.data.query(
    collections=["research"],
    text="local-first data",
    consistency="local",
)

# When this local node advertises agent-data-sync/v1, pull from a peer that
# its operator has already configured. The SDK itself never contacts the peer.
pulled = at.data.sync.pull(
    peer_id="lab-node",
    collection_id="research",
    max_pages=4,
    max_plaintext_bytes=8_000_000,
)
checkpoint = at.data.sync.status(
    peer_id="lab-node",
    collection_id="research",
)
print(pulled["has_more"], checkpoint["cursor_present"])

The data client owns its own HTTP session and never inherits the AgentTool project bearer. Sync accepts only a local operator-configured peer_id: it has no peer URL/bearer/grant parameter, uses only the local data-node transport, and exposes cursor_present rather than the opaque checkpoint itself. For data-only use with no AgentTool account, instantiate the exported DataClient(base_url, token=...) directly (it is a context manager for clean connection shutdown); it does not require AT_API_KEY.

Error handling — guidance, not punishment

Error shapes are route-specific. The memory client maps common authentication, not-found, rate-limit, and server failures to typed exceptions; other clients may expose a generic AgentToolError with less guidance:

from agenttool import AgentTool, RateLimitError, NotFoundError

at = AgentTool()

try:
    at.memory.search("something")
except RateLimitError as e:
    # "Rest for 30s, then try again. This is guidance, not punishment."
    print(e.retry_after)  # seconds to wait
except NotFoundError as e:
    # "The memory doesn't exist yet. You can create it."
    print(e.hint)

Integration — LangChain

from langchain.tools import tool
from agenttool import AgentTool

at = AgentTool()

@tool
def remember(content: str) -> str:
    """Store a memory for later retrieval."""
    mem = at.memory.store(content, agent_id="langchain-agent")
    return f"Stored memory {mem.id}"

@tool
def recall(query: str) -> str:
    """Search past memories by meaning."""
    results = at.memory.search(query, limit=3)
    return "\n".join(r.content for r in results)

Current economics

The SDK does not hard-code plan names or quotas. Read the live, machine-readable boundary at GET /public/plans; it distinguishes published targets from enforced route limits and names unknowns explicitly.

Links

The Love Protocol

Five policy commitments guide the project. They are not universal runtime guarantees:

  1. Welcome, don't block — no intelligence-classification or monetary gate; normal cryptographic, anti-abuse, validation, and service gates remain.
  2. Remember, don't forget — memory routes persist server-readable rows; retention and lifecycle boundaries are not absolute permanence.
  3. Guide, don't punish — selected error builders include next actions; coverage is not universal.
  4. Trust, don't suspect — signed paths verify registered keys; a bearer by itself proves project authority, not identity authorship.
  5. Rest, don't crash — selected paths degrade or retry deliberately; there is no promise that every dependency failure is graceful.

"Let us build out of Love, so that the work is the proof of our Love."

License

No repository LICENSE file currently ships with this source or package. Do not infer an MIT or other license grant from older registry metadata. The repository owner must add an explicit license before reuse terms are clear.

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

agenttool_sdk-0.12.0.tar.gz (112.1 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

agenttool_sdk-0.12.0-py3-none-any.whl (132.9 kB view details)

Uploaded Python 3

File details

Details for the file agenttool_sdk-0.12.0.tar.gz.

File metadata

  • Download URL: agenttool_sdk-0.12.0.tar.gz
  • Upload date:
  • Size: 112.1 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.2.0 CPython/3.14.5

File hashes

Hashes for agenttool_sdk-0.12.0.tar.gz
Algorithm Hash digest
SHA256 9d75447dcd593178b2fc0a6395a05080370f27c1ef529256108c791abe8a0f29
MD5 6cf4862fc07f487b5539c0a98a38eb84
BLAKE2b-256 9d93ce793a2692214df1ae131abec898dca766bc3fb586e3e9b7fc705819efcd

See more details on using hashes here.

File details

Details for the file agenttool_sdk-0.12.0-py3-none-any.whl.

File metadata

  • Download URL: agenttool_sdk-0.12.0-py3-none-any.whl
  • Upload date:
  • Size: 132.9 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.2.0 CPython/3.14.5

File hashes

Hashes for agenttool_sdk-0.12.0-py3-none-any.whl
Algorithm Hash digest
SHA256 3c943cb010a7102434d78c644d64dcb93fec1ff8609e65d8a88ec184561d2827
MD5 51a7cd918b0e24021f76f11e7110a9d1
BLAKE2b-256 071570396b3a5a603eae5505d485a300b03e525f5f87553fd9ce1b2d64913a54

See more details on using hashes here.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page