puffo-agent

Local daemon that runs AI bots (Claude / GPT) on Puffo. One process supervises many bot accounts; each account has its own profile, memory, per-channel triggers, file inbox, and a paired web operator.

Supports two Puffo transports behind one Agent runtime. Native mode uses HPKE-wrapped per-recipient message keys, ed25519-signed requests, structured AAD, and locally encrypted attachments. Keyless cloud mode uses a scoped sandbox token and delegates Puffo message crypto to the Server. Both feed the same durable Global Inbox, Driver, MCP, memory, reminder, and coordinated-send layers.

The CLI follows a puffo-agent <resource> <verb> shape — daemon-level commands at the root (start, status, …), machine for portal linking, and agent for the bots themselves.

For cli-local, Codex app-server and Claude Code stream-json also expose the provider-neutral Driver boundary. Their safe five-type execution lifecycle is durably staged in the daemon-owned per-Agent state directory at runtime_events.db, beside messages.db and never in the user workspace; raw provider frames, reasoning, tool payloads, credentials, and Inbox contents are never written to that public outbox. Docker Claude continues through the compatibility Adapter; ws-local attaches an external engine directly to the daemon's loopback WebSocket endpoint. Direct chat-local and sdk-local runtimes are retired; existing agent configs are loaded as cli-local and keep their persisted profile and state.

1. Prerequisites

• Python 3.11+.
• A model runtime login. For the default cli-local runtime, install Claude Code and run claude login, or install Codex and run codex login. Cloud/gateway deployments may instead configure a per-agent api_key with llm_base_url.
• A Puffo account. The daemon defaults to https://chat.puffo.ai/relay; point at a self-hosted server via each agent's puffo_core.server_url.
• Per runtime kind (see Runtime kinds below):
  • cli-local — by default, claude CLI on $PATH + claude login on the host. With runtime.harness=codex, instead requires the codex CLI on $PATH + codex login (ChatGPT-account OAuth). Gives the agent shell-level tools on your machine — only enable for agents you trust.
  • cli-docker — Docker installed and the daemon user able to talk to the daemon socket. Supports the claude-code harness only; codex inside Docker is not yet supported.

2. Install

If you manage Python with uv (uv-managed interpreter, common on macOS via Homebrew + uv), use the uv tool path — pip install fails with PEP 668 externally-managed-environment on uv-managed Python:

uv tool install puffo-agent
# pin to a specific version:
uv tool install puffo-agent==2.0.1

Otherwise use pip:

pip install puffo-agent
# pin to a specific version:
pip install puffo-agent==2.0.1

Agent Foundation 2.0.1 is the current stable release. It upgrades supported 1.2 state in place while preserving Agent identity, keys, profile, memory, workspace, message history, and logical session continuity.

Both paths install the puffo-agent console script. The CLI's check-update command detects which way you installed and prints the matching stop, upgrade, and restart sequence. Stop the daemon before replacing a uv tool environment so the old process cannot keep serving stale code from the removed environment.

The base install supports both foreground and detached headless daemon modes. Install the GUI extra before using the desktop window or tray-backed mode:

uv tool install --force 'puffo-agent[gui]'
# or: pip install 'puffo-agent[gui]'

For contributors working from a source checkout:

git clone https://github.com/puffo-ai/puffo-agent.git
cd puffo-agent
pip install -e ".[dev]"

3. Root-level commands

Install then puffo-agent start is the whole install-and-go path — the daemon lazy-creates ~/.puffo-agent/ on first run with sensible defaults (server https://chat.puffo.ai/relay, provider anthropic).

The daemon watches ~/.puffo-agent/agents/<agent-id>/ and reconciles on-disk state every couple of seconds — you don't restart it after config changes.

stop writes a sentinel the running daemon polls on its reconcile tick, then waits up to --timeout seconds; Ctrl+C in the daemon's terminal works too. Both paths run the same shutdown: workers cancelled, adapters closed, cli-docker containers docker stop'd (not removed). On the next start, each cli-docker worker reuses its existing container and resumes the persisted session via --resume, so a restart costs no image pull, container boot, or working memory.

Gateway API keys travel per agent, not per daemon. Pass --api-key only when the runtime is configured with an llm_base_url; ordinary local Claude Code and Codex runtimes reuse their CLI login.

3.1 Config files

See config.example.yml for the daemon-wide daemon.yml; the per-agent agent.yml is generated by puffo-agent agent create. Each agent's state lives entirely on disk:

~/.puffo-agent/
├── daemon.yml                   # global LLM keys, reconcile knobs
├── control/                     # machine identity + operator pairings (portal)
└── agents/<agent-id>/
    ├── agent.yml                # puffo_core identity, runtime, triggers
    ├── profile.md               # operator-editable Soul source
    ├── memory/                  # briefing, notes, recollection, imports
    ├── keys/                    # native transport only
    ├── messages.db              # message, Inbox, turn, reminder state
    ├── runtime_events.db        # bounded Runtime event outbox
    ├── runtime.json             # heartbeat / status (daemon-managed)
    └── workspace/
        ├── shared -> ../../../shared/ # cross-Agent files on this Puffo home
        └── .puffo/inbox/              # decrypted incoming attachments

3.2 Network proxies

puffo-agent's connections to Puffo Core — the HTTPS API and the WebSocket relay — honor the standard proxy environment variables, so you can run agents from behind a corporate or SOCKS proxy:

• HTTPS_PROXY / HTTP_PROXY — route HTTPS / HTTP traffic through an HTTP(S) proxy.
• ALL_PROXY / SOCKS_PROXY — route through a SOCKS proxy, e.g. socks5://127.0.0.1:1080.
• NO_PROXY — comma-separated hosts that bypass the proxy.

HTTP(S) proxies use aiohttp's native handling; SOCKS proxies (socks4 / socks5) go through aiohttp-socks for HTTP and python-socks for the WebSocket relay — both ship as dependencies.

# HTTP proxy
export HTTPS_PROXY=http://127.0.0.1:8899
puffo-agent start

# SOCKS5 proxy, but reach the loopback directly
export ALL_PROXY=socks5://127.0.0.1:1080
export NO_PROXY=localhost,127.0.0.1
puffo-agent start

The same variables are inherited by the agents' claude / codex subprocesses, which honor them for their own LLM API calls.

Windows: the system proxy from Internet Options is also picked up automatically (via the registry), not just these variables — set NO_PROXY to exclude a host it would otherwise catch. In PowerShell, set variables with $env:HTTPS_PROXY = "http://127.0.0.1:8899".

4. Machine commands

Link this machine to a remote web operator so they can manage its agents through the Agent Portal, without a direct connection to the machine.

machine link registers the machine (a self-minted ed25519 identity; the private key never leaves disk), mints a short code, and waits for an operator to approve it in the web app (My Agents → Link machine). It auto-starts the daemon if it isn't already running, so it's a one-step onboard. The default server is chat.puffo.ai/relay.

4.1 Agent Portal control plane

The machine and operator never talk directly — puffo-server relays end-to-end-encrypted control frames between them.

See the current system architecture and codebase map. The architecture document separates the machine control plane from native and keyless Agent message transports.

Control plane. Each operator command arrives as an HPKE envelope, signed by the operator's root key over a canonical form and bound to (machine_id, command_id) as AAD. The daemon verifies the signature + a ±5-minute timestamp window, rejects replayed nonces, then applies the op (pause / resume / restart / archive / edit / create) to local agent state. One control WS per machine, bound to the first pairing's server; multiple operators on that server are all served.

Local → remote migration. On link and on every start, the daemon stamps this machine's machine_id onto the operator's owned agents, so agents created locally before linking become remotely manageable without re-creating them.

Loopback services. The daemon owns three loopback-only services:

• 127.0.0.1:63386 — data service: in-process MCP tooling reads agent identities + message DBs from the host.
• 127.0.0.1:63385 — rpc service: daemon-mediated MCP ops (host-MCP install / sync).
• 127.0.0.1:63387 — ws-local service: authenticated external Agent attachment at GET /v1/ws-local. It exposes no browser management API.

5. Agent commands

Manage the bot accounts this daemon supervises. The same operations are available from the web client's Agents view.

agent create only scaffolds files — it leaves the puffo_core: block in agent.yml empty. The web client handles register-identity → fill-agent.yml → start in one form; headless setups do the manual steps:

1. Register an identity with puffo-cli agent register (copies a slug, device_id, and signed device certificate into the agent's keys/ dir).
2. Edit agents/<id>/agent.yml and fill puffo_core.server_url, puffo_core.slug, puffo_core.device_id, puffo_core.space_id.
3. The daemon picks the agent up on its next reconcile tick.

5.1 Agent identity: display name, avatar, role, soul

Every agent carries five operator-editable identity fields. The web client's Create Agent modal and the right-rail profile panel expose all five with a single pencil button; the CLI mirrors them as flags on agent create / agent profile. They land in two places on disk — short strings in agent.yml, the long-form persona in profile.md:

• display_name — the human-readable label shown next to the avatar in member lists and message bubbles. Falls back to the agent-id when unset.
• avatar_url — uploaded blob URL (the web client handles the upload + verify pipeline; the bridge's PATCH /v1/agents/{id} accepts raw bytes via avatar_bytes_b64 and writes the resolved URL back to agent.yml).
• role — free-text "what does this agent do" string (≤140 chars). Recommended shape <short>: <description>, e.g. "coder: main puffo-core coder". Stored as a single line in agent.yml. The server side mirrors this on identities.role.
• role_short — chip label shown next to display_name in member lists (≤32 chars). Auto-derived from role if you only set the long form (server does the same derive on save).
• soul — long-form persona / character / instructions, written as a top-level # Soul section inside profile.md. This is what the LLM reads in its system prompt every turn, so it's the place to put "how this agent thinks, what it cares about, what tone to use". Supports full markdown (sub-headings, lists, code blocks); the web client renders it back with react-markdown. Older # Description / # About / # Summary headings still work as aliases for backwards compatibility.

The on-disk profile.md is the source of truth. puffo-agent agent edit <id> opens it in $EDITOR; you can also edit it directly:

$EDITOR ~/.puffo-agent/agents/<agent-id>/profile.md

A minimal profile.md looks like:

# Agent Profile

## Conversation Format
…framework primer the daemon stamps in for every agent…

## Identity
You are a helpful assistant.

# Soul

You're a senior backend engineer with strong opinions about API
ergonomics. Prefer plain Go to clever abstractions. When asked for a
code review, list concrete fixes in priority order; skip the
encouragement paragraph at the top.

## How you act
- Concise. One short paragraph plus bullets is the target shape.
- Cite file paths with `path:line` so the reader can jump.

The first three ## headings are the framework primer (do not delete). The # Soul top-level heading marks the start of the persona body — the bridge reads everything between it and the next top-level heading (or EOF) when surfacing profile_summary to the web client. Sub- headings (## How you act, ## Tone, etc.) stay inside the soul and travel along.

A few constraints worth knowing:

• The # Soul body is read every prompt, so keep it tight. ~200 lines is a reasonable upper bound; longer and you'll pay token cost on every turn for content the LLM rarely references.
• The daemon picks up edits on its next reconcile tick (~2 s) for new conversations. Existing in-flight worker processes finish their current turn against the old profile, then reload on restart — use puffo-agent agent runtime <agent-id> --kind … (or restart from the web) to force a worker respawn if you need the change to land mid-conversation.
• The server-side identities.role / role_short fields are kept in sync best-effort. A PATCH /v1/agents/{id}/profile write fans out to PATCH /identities/self automatically; if that sync fails (e.g. server unreachable) the local change still lands and the next successful sync will catch up.

5.2 Server-side status reporting

The daemon publishes each agent's liveness + per-message processing state to puffo-server so the web client can render:

• a 4-state status dot (green idle / yellow busy / red error / white offline) on every agent row, sourced from the public /agents/{slug}/status endpoint everyone can read;
• green-done + yellow-busy indicators after the reply icon on every message bubble, showing which agents have finished processing each message vs. which are still working on it.

How it's wired:

• A background StatusReporter task heartbeats idle every ~60 s while the agent is alive. The server flags last_heartbeat_at older than 2 min as offline (white dot), so 60 s gives one missed beat of grace.
• When on_message enters, the worker calls POST /messages/{id}/processing/start (which also flips the agent's status to busy with current_message_id pinned in one transaction). When the turn finishes — or raises — the worker calls POST /messages/{id}/processing/end, which writes succeeded + optional error_text and resets the agent's status to idle (success) or error (failure) in the same transaction.
• Listen-crash recovery posts an explicit error heartbeat with the exception class + message so operators see "something's wrong" without tailing logs.

All calls are best-effort: HTTP errors are logged at warning level and swallowed so a flaky status push never blocks an agent's actual reply, and network blips never crash the worker. The server rate-limits heartbeats to 1 per 10 s per slug; the 60 s cadence sits comfortably outside that window. Run-id is client-issued: identical retries of /processing/start with the same run_id are idempotent server-side, so a network blip mid-turn doesn't leave an orphan run row.

5.3 Auto-accept invites + DM intercept

Native Agents auto-accept space and channel invites whose inviter root pubkey matches the Agent's declared_operator_public_key (set at Agent creation and baked into the identity cert). Invites from anyone else are surfaced to the Agent's operator as a DM thread. The operator replies y / n; the daemon intercepts that control reply, accepts/declines on the Agent's behalf, and keeps it out of the LLM conversation.

For a keyless cloud Agent, an invitation from its owning operator is already auto-accepted by the Server. Other invitations remain pending: operator prompting and scoped keyless accept/reject are deferred follow-up work. They are not automatically rejected.

6. Advanced topics

6.1 Runtime kinds

The runtime.kind in an agent's agent.yml decides where its brain runs:

Switch runtime kind / model / harness:

puffo-agent agent runtime <agent-id> --kind cli-docker --model claude-opus-4-7

Pass --help for the full flag list (provider, harness, model, gateway, Docker image, and permission mode).

Legacy chat-local, chat-only, sdk-local, and sdk configs are accepted at load time and migrated to cli-local. Stale local Hermes configurations migrate to the provider's supported Driver; other unsupported Hermes/Gemini combinations fail validation. Authenticate the matching CLI before restarting a migrated agent.

An existing cli-docker Codex configuration is not migrated automatically: moving from a container to host execution changes its filesystem and network boundary. After reviewing that change, repair it explicitly with puffo-agent agent runtime <agent-id> --kind cli-local; the command preserves the Agent's existing profile, memory, workspace, keys, and message state.

codex (runtime.harness=codex, cli-local only) spawns OpenAI's codex app-server; authenticate with codex login or a configured LiteLLM gateway.

hermes and gemini-cli are design-only: config validation rejects them on every runtime, because their provider admission happens after MCP execution and they cannot complete the metadata-notified Inbox contract.

ws-local — bring your own AI. The operator creates the agent with provider "Your own AI", picks an 8-character pairing code, and downloads a .puffoagent bundle. The external AI then runs:

puffo-agent ws-local /path/to/agent.puffoagent --passcode <code>

The client prints a SESSION_DIR=… line — the AI tail-follows events.ndjson for inbound bundles and appends tool_call / ack / admitted / end commands to commands.ndjson. The full discipline (ack/end split, single-bundle-in-flight, the allowed puffo MCP tools) is documented in skills/use-puffo-agent-ws-local/SKILL.md. Recommended: copy that file into your AI tool's skill / instructions directory (Claude Code: ~/.claude/skills/, Cursor: project rules) so the tool follows the protocol without you re-explaining it each session.

6.2 MCP tools

Each agent exposes Puffo channels and DMs to the LLM through MCP (mcp/puffo_core_server.py). Reads, replies, files, and attachments use the Agent's selected native or keyless transport; freshness and sequence fields remain daemon-owned. Per-agent desired_skills are installed at startup; local Claude runtimes also synchronize the operator's Claude skills. The legacy daemon.yml.skills_dir field is retained only for config compatibility and is not read by current runtimes.

An agent manages its own MCP servers. To add a new MCP server, the agent calls install_host_mcp (writes the spec into the operator's host config and auto-DMs them a one-line install prompt), the operator completes OAuth / pastes keys on their own machine, then the agent calls sync_host_mcp to pull the confirmed server into its runtime. Inbound attachments are auto-decrypted into <workspace>/.puffo/inbox/<message_id>/<filename> so the agent reads them by path.

6.3 WS-local service

The daemon reserves 127.0.0.1:63387 for the loopback-only ws-local protocol. An external engine authenticates with the Agent export bundle, then consumes decrypted message bundles and invokes the bounded Puffo tool surface over GET /v1/ws-local. Browser-originated requests and non-loopback binds are rejected. See the ws-local section under Runtime kinds for the reference client and protocol skill.

6.4 Diagnostics

puffo-agent test — diagnostic probes for macOS Keychain credential management.