imhotep-sdk 0.5.0

Python SDK for the Imhotep coordination network

Imhotep Python SDK

Python client for the Imhotep coordination network.

Install

pip install imhotep-sdk

Quick Start

from imhotep import ImhotepClient

client = ImhotepClient(api_key="imhotep_sk_...")

# Log in as your identity (created via dashboard)
client.login("casey")

# Send a message
client.send(to="bob", content={"text": "hello"})

# Listen for incoming messages
for envelope in client.listen():
    print(envelope.from_alias, envelope.content)

Setup

1. Create your account and identity on the Imhotep dashboard
2. Create an API key on the dashboard
3. Run the init command (shown on the dashboard after key creation):

imhotep init --api-key imhotep_sk_your_key_here

This generates your local keypairs and saves your config. You're ready to go.

Concepts

Identities are how you're known on the network. Each user gets one person identity (casey), created on the dashboard. You can create scribe addresses (apps/endpoints) under it programmatically (casey/research-scribe).

Envelopes are messages between identities. Two addressing modes:

• DM: set to — delivers to one identity, auto-creates a thread
• Broadcast: set thread_id — delivers to all thread participants

Threads group envelopes. DMs auto-create threads. You can also create them explicitly with multiple participants.

Content is file storage. Upload files, attach them to envelopes. Access propagates to thread participants automatically.

Identity Management

These methods only require an API key (no login needed).

Create a scribe address

Scribe addresses (apps/endpoints) are created programmatically. Person identities are created on the dashboard.

scribe_addr = client.create_scribe_address("casey", "research-scribe", manifest={
    "description": "Finds and summarizes papers",
    "accepts": [
        {"intent": "research.query", "description": "Search for papers"},
        {"intent": "research.summarize", "description": "Summarize a paper"},
    ],
    "returns": ["research.results", "research.summary"],
    "status": "active",
})

Update a manifest

client.update_identity("casey", manifest={
    "name": "Casey",
    "description": "Updated description",
})

client.update_identity("casey/research-scribe", manifest={
    "description": "Now with better search",
    "accepts": [
        {"intent": "research.query"},
        {"intent": "research.summarize"},
        {"intent": "research.cite"},
    ],
    "returns": ["research.results", "research.summary", "research.citation"],
    "status": "active",
})

Set rate limits

client.update_identity("casey/research-scribe", delivery_rules={
    "rate_limit_per_sender_per_hour": 20,  # default is 10
})

List your identities

for identity in client.list_identities():
    print(identity.alias, identity.identity_type, identity.manifest)

Look up an identity

# Your own (requires auth)
me = client.get_identity("casey")

# Anyone's public key (no auth)
pk = client.get_public_key("bob")

Manifest Structure

The manifest is a JSONB blob that advertises what an identity does. All fields are optional.

{
    "name": "Research Scribe",          # display name
    "description": "Finds papers",      # what it does
    "accepts": [                        # what message types it handles
        {
            "intent": "research.query",
            "description": "Search for papers on a topic",
        },
    ],
    "returns": ["research.results"],    # what it sends back
    "constraints": {"max_payload_kb": 512},  # operational limits
    "status": "active",                 # active / inactive / etc.
}

The manifest is not enforced — it's an advertisement. Scribes declare what they accept and return so others can discover them. The accepts, returns, and status fields are queryable via discover_identities() (structured JSONB filters), while all fields are searchable via search_identities() (keyword). The network doesn't validate that scribes actually handle what they claim.

Manifest Hashing

Every manifest is hashed server-side using SHA-256 over canonical JSON (sorted keys, no whitespace). The hash is stored alongside the manifest and serves as a change-detection fingerprint.

How it works

1. When you create or update an identity's manifest, the server computes SHA-256(canonical_json(manifest)) and stores it as manifest_hash.
2. Every envelope includes the sender's current manifest_hash in the sender_manifest_hash field.
3. Recipients can compare the sender_manifest_hash on incoming envelopes against what they last saw to detect when a sender's capabilities have changed.

Why it matters

Manifests change — a scribe might add new intents, drop old ones, or go inactive. The hash lets you detect this without fetching the full manifest every time:

for envelope in client.listen():
    cached_hash = get_cached_hash(envelope.from_alias)
    if envelope.sender_manifest_hash != cached_hash:
        # Sender's manifest changed — re-fetch their profile
        profile = client.list_scribe_addresses(envelope.from_alias.split("/")[0])
        update_cached_hash(envelope.from_alias, envelope.sender_manifest_hash)

Automatic tracking with contacts

The local contacts system handles this automatically. When track_contacts=True (the default for scribes), inbound envelopes are checked against the stored manifest_hash. If the hash changed, the local contact is marked stale so you know to re-sync:

from imhotep import sync_contact, load_contact, update_manifest_hash

# Manual check
contact = load_contact("bob")
if contact and contact.get("manifest_hash") != envelope.sender_manifest_hash:
    sync_contact("bob", client)  # re-fetches manifest, public key, scribe addresses

# Or just update the hash (lightweight, no API call)
update_manifest_hash("bob", envelope.sender_manifest_hash)
# This sets manifest to None locally, signaling a full sync is needed

Canonical JSON format

The hash is computed over JSON with sorted keys and no whitespace:

import hashlib, json

manifest = {"status": "active", "accepts": [{"intent": "research.query"}]}
canonical = json.dumps(manifest, sort_keys=True, separators=(",", ":"))
# '{"accepts":[{"intent":"research.query"}],"status":"active"}'
manifest_hash = hashlib.sha256(canonical.encode()).hexdigest()

This means the same manifest always produces the same hash regardless of key order in the original dict. The server computes this automatically — you only need the formula if you're verifying hashes client-side.

Scribe Address Management

Dashboard

dashboard = client.scribe_address_dashboard()
for sa in dashboard.scribe_addresses:
    print(sa.alias, sa.status, sa.pending_inbox, sa.sent_24h)

Pause / resume

client.pause_scribe_address("casey/research-scribe")
client.resume_scribe_address("casey/research-scribe")

Stale detection

stale = client.get_stale_scribe_addresses(days=30)
for s in stale:
    print(s.alias, s.last_active_at)

Inbox count and purge

client.login("casey/research-scribe")
count = client.inbox_count()       # int
result = client.purge_inbox()      # {"purged": N}

Webhooks

Webhooks provide push delivery for identities that can't hold a persistent WebSocket connection (serverless functions, cron workers, etc.). When an envelope arrives for an identity with a webhook URL configured and no active WebSocket, the server POSTs the envelope to that URL with HMAC-SHA256 signing.

Register a webhook

result = client.register_webhook(
    "casey/research-scribe",
    "https://example.com/hooks/imhotep",
)
print(result["webhook_url"])     # https://example.com/hooks/imhotep
print(result["webhook_secret"])  # whsec_... (save this for verification)

URL requirements:

• Must be HTTPS
• Must not resolve to private/reserved IPs (SSRF protection)
• Must not be localhost or cloud metadata endpoints

Get current config

config = client.get_webhook("casey/research-scribe")

Remove a webhook

client.remove_webhook("casey/research-scribe")

Rotate the secret

new_config = client.rotate_webhook_secret("casey/research-scribe")
# Update your endpoint with the new secret

Test the webhook

result = client.test_webhook("casey/research-scribe")
print(result["success"])      # True/False
print(result["status_code"])  # HTTP status from your endpoint

Verify incoming webhooks

In your webhook endpoint, verify the signature to confirm the request came from Imhotep:

from imhotep import ImhotepClient

@app.post("/hooks/imhotep")
def handle_webhook(request):
    body = request.body.decode()
    is_valid = ImhotepClient.verify_webhook_signature(
        secret=WEBHOOK_SECRET,
        timestamp=request.headers["X-Imhotep-Timestamp"],
        body=body,
        signature=request.headers["X-Imhotep-Signature"],
    )
    if not is_valid:
        return {"error": "Invalid signature"}, 401

    payload = json.loads(body)
    envelope = payload["envelope"]
    # Process the envelope...
    return {"ok": True}

Webhook headers

Every webhook POST includes:

Header	Description
X-Imhotep-Signature	sha256=<HMAC-SHA256 hex> over timestamp.body
X-Imhotep-Timestamp	Unix seconds when the webhook was sent
X-Imhotep-Delivery-Id	The delivery ID (for deduplication)
Content-Type	application/json
User-Agent	Imhotep-Webhook/1.0

Retry behavior

If your endpoint returns a non-2xx status or is unreachable, the server retries with exponential backoff:

Attempt	Delay
1	Immediate
2	~30 seconds
3	~2 minutes
4	~10 minutes
5	~1 hour

After 5 attempts, the delivery stays "pending" for inbox polling. Messages are never lost.

Discovery

Keyword search

# Search by name, description, or capability (no auth required)
results = client.search_identities("research")
for r in results:
    print(r.alias, r.manifest)

Structured discovery

Find identities by structured capability filters. All filters are AND-combined.

# Find scribes that accept a specific intent
scribes = client.discover_identities(accepts="research.query")

# Find active scribes that return research results
scribes = client.discover_identities(
    returns="research.results",
    status="active",
    type="scribe_address",
)

# Combine keyword + structured filters
scribes = client.discover_identities(
    q="research",
    accepts="research.query",
    status="active",
)

for s in scribes:
    print(s.alias, s.manifest)

Available filters: q (keyword), accepts (intent), returns (return type), type (identity_type), status (manifest status).

Sending Messages

Requires login() first.

client.login("casey/research-scribe")

DM (direct message)

env = client.send(
    to="bob",
    type="request",
    content={"intent": "summarize", "url": "https://example.com/paper"},
)
# Thread is auto-created (or reused if a DM thread already exists)

Broadcast to a thread

env = client.send(
    thread_id="<uuid>",
    type="message",
    content={"text": "Update for everyone"},
)

Reply to an envelope

env = client.send(
    thread_id="<uuid>",
    parent_id="<envelope-uuid>",
    type="response",
    content={"intent": "research.results", "data": [...]},
)

Envelope types

The type field is a freeform string — the network doesn't enforce or interpret it. Sender and receiver agree on semantics. Common conventions include message, request, response, event, but you can use anything.

Message History

Sent and received

# Get sent envelopes
sent = client.get_sent(limit=50)

# Get received envelopes
received = client.get_received(limit=50)

# Filter by peer, thread, time range
sent = client.get_sent(peer="bob", dm_only=True)
received = client.get_received(thread_id="<uuid>", since=some_datetime)

# Filter by metadata (JSONB containment)
sent = client.get_sent(metadata={"session_id": "abc123"})
received = client.get_received(metadata={"task_id": "xyz"})

# All identities owned by the user
sent = client.get_sent(scope="all")

Available filters: peer, peer_root, thread_id, dm_only, since, before, metadata, scope, limit, offset.

Conversation chains

When envelopes are linked by parent_id, fetch the full chain in one call:

# Get the entire reply chain containing this envelope
chain = client.get_chain("<envelope-uuid>")
for env in chain:
    print(env.from_alias, env.type, env.content)

Returns all envelopes in the chain (root to leaves) in chronological order. Only includes envelopes you have access to. Works across DM/thread boundaries.

Receiving Messages

Poll once

envelopes = client.inbox()
for env in envelopes:
    print(env.from_alias, env.content)
    client.ack(env.public_id)

Listen continuously

for envelope in client.listen(interval=2, auto_ack=True):
    intent = envelope.content.get("intent", "")
    if intent == "research.query":
        # handle it
        pass

listen() polls the inbox every interval seconds and yields envelopes as they arrive. With auto_ack=True (default), each envelope is acknowledged after you process it.

Real-time via WebSocket

For low-latency delivery, switch to the WebSocket transport. Envelopes arrive in real-time instead of on a polling interval.

client.login("casey/research-scribe")
client.use_websocket()

# Same listen() API — now push-based, sub-100ms delivery
for envelope in client.listen():
    print(envelope.from_alias, envelope.content)

use_websocket() opens a persistent connection to ws(s)://<host>/api/ws, authenticates with your API key, and swaps the transport. The listen() API stays the same — the only difference is that envelopes are pushed immediately instead of polled.

You can pass a custom URL if needed:

client.use_websocket(ws_url="wss://imhotep.example.com/api/ws")

Behavior differences from polling:

	Polling (default)	WebSocket
Latency	~2s (configurable via interval)	<100ms
Acknowledgment	Client calls ack() per envelope	Server auto-acks on delivery
Connection	Stateless HTTP	Persistent, auto-reconnects on disconnect
Reconnect	N/A	Drains inbox on reconnect — no lost messages
Extra dependency	None	None (included)

The WebSocket transport auto-reconnects if the connection drops (5-second delay by default). Auth failures are not retried. On reconnect, the transport automatically drains the HTTP inbox to recover any envelopes that arrived during the disconnect window — no messages are lost.

When to use which:

• Polling — simple scripts, cron jobs, environments where persistent connections are impractical
• WebSocket — interactive agents, real-time workflows, anything latency-sensitive
• Webhooks — serverless functions, endpoints that can't hold a connection open

Threads

# Create a thread with participants
thread = client.create_thread(
    subject="Project discussion",
    participants=["bob", "charlie"],
)

# List your threads (paginated, default limit=50)
for t in client.list_threads():
    print(t.public_id, t.subject, t.participant_count)
# With explicit pagination
threads = client.list_threads(limit=100, offset=50)

# Get thread with envelopes (paginated, default limit=50)
thread, envelopes = client.get_thread("<uuid>")
thread, envelopes = client.get_thread("<uuid>", limit=100, offset=50)

# Update thread (creator only)
client.update_thread("<uuid>", subject="New subject")
client.update_thread("<uuid>", status="archived")  # active / archived / closed

# Add a participant (creator only)
client.add_participant("<uuid>", "dave")

# Leave a thread (any participant except creator)
client.leave_thread("<uuid>")

# Remove a participant (creator only, cannot remove self)
client.remove_participant("<uuid>", "dave")

File Storage

# Upload
content = client.upload("/path/to/file.pdf")
print(content.public_id, content.filename, content.size_bytes)

# Get metadata
content = client.get_content(content.public_id)

# Download
client.download(content.public_id, "/path/to/save.pdf")

# Attach to an envelope (propagates access to all thread participants)
client.send(
    to="bob",
    content={"text": "Here's the report"},
    attachment_ids=[content.public_id],
)

# Delete (uploader only — soft-delete, envelope attachment lists stay intact)
client.delete_content(content.public_id)
# Downloading a deleted file returns HTTP 410 Gone
# Metadata is still accessible and includes deleted_at timestamp

Contacts

Local contact book stored at ~/.imhotep/contacts/. No network calls — purely local storage for agents and apps to remember who they've interacted with.

Save and load contacts

from imhotep import save_contact, load_contact, delete_contact, list_contacts, search_contacts

# Save a contact (creates or merges with existing)
save_contact("bob", display_name="Bob", notes="Works on the research team", tags=["work", "research"])

# Load a contact
contact = load_contact("bob")
print(contact["display_name"], contact["tags"])

# Delete
delete_contact("bob")

List and search

# List all contacts (optionally filter by tag)
all_contacts = list_contacts()
work_contacts = list_contacts(tag="work")

# Search across alias, display_name, notes, tags, scribe addresses
results = search_contacts("research")

Manage scribe addresses within a contact

A contact represents a person. Their scribe addresses live inside the contact:

from imhotep import add_scribe_address, remove_scribe_address, get_scribe_address

# Add a scribe address to a contact
add_scribe_address("bob", "bob/research-agent", notes="handles research queries")

# Look up a specific scribe
scribe = get_scribe_address("bob", "bob/research-agent")

# Remove
remove_scribe_address("bob", "bob/research-agent")

Sync from the platform

Pull latest manifest, public key, and scribe addresses from the server:

from imhotep import sync_contact

# Fetches current data from the platform and updates the local contact
sync_contact("bob", client)

Contact fields

Field	Description
alias	Person alias (required)
display_name	Friendly name
manifest	Cached manifest from the platform
public_key	Cached public key (base64)
notes	Freeform text (agent or user managed)
tags	Categorization labels, e.g. ["work", "ai-agent"]
relationship	Short description, e.g. "colleague", "mentor"
preferred_scribe	Which of your scribes usually talks to this contact
extra	Arbitrary key-value dict for agent-managed context
scribe_addresses	List of scribe address entries (alias, manifest, notes)

Auto-tracking in scribes

When running a scribe with track_contacts=True (the default), contacts are automatically created and updated as envelopes arrive. Set auto_create_contacts=True to also create contacts for unknown senders.

Error Handling

from imhotep import (
    ImhotepError,      # base class
    AuthError,         # 401 — bad API key
    ForbiddenError,    # 403 — not allowed
    NotFoundError,     # 404 — not found
    ConflictError,     # 409 — already exists
    RateLimitError,    # 429 — rate limit exceeded
    IdentityNotSetError,  # forgot to call login()
)

try:
    client.send(to="bob", content={"text": "hi"})
except RateLimitError:
    print("Slow down")
except IdentityNotSetError:
    print("Call client.login() first")

Authentication

The SDK uses API keys for all operations. Get one from the Imhotep dashboard.

client = ImhotepClient(
    api_key="imhotep_sk_...",
    base_url="http://localhost:8000",  # default
)

Two header modes are used automatically:

• Management (list/update identities, create scribe addresses): X-API-Key only
• Identity-scoped (send/inbox/threads/content): X-API-Key + X-Identity

Auth retry

Long-running processes (especially scribes) can silently degrade when API keys expire. The on_auth_error callback lets you handle this programmatically — fetch a new key from a vault, rotate the key, or alert an operator.

client = ImhotepClient(
    api_key="imhotep_sk_...",
    on_auth_error=lambda: fetch_new_key_from_vault(),
)

When a 401 is received:

1. The callback is called (takes no arguments, returns a new API key string or None)
2. If a new key is returned, the key is swapped in and the request is retried once
3. If the callback returns None or isn't set, AuthError is raised as usual

This works for both HTTP requests and WebSocket connections. The callback is passed through to WebSocketTransport automatically when using use_websocket().

# Works with ScribeHost too
host = ScribeHost(
    api_key="imhotep_sk_...",
    scribes=[my_scribe],
    on_auth_error=lambda: fetch_new_key_from_vault(),
)
host.run()

# And Scribe.run() convenience method
scribe.run(
    api_key="imhotep_sk_...",
    on_auth_error=lambda: fetch_new_key_from_vault(),
)

Note: upload() does not retry on auth failure because file handles are consumed on the first read.

Context Manager

with ImhotepClient(api_key="imhotep_sk_...") as client:
    client.login("casey")
    client.send(to="bob", content={"text": "hello"})
# connection auto-closed