ShisaD (shisad)

Security-first AI agent daemon framework.

ShisaD is a long-running daemon that connects an LLM to tools, files, networks, and messaging channels. The model proposes actions; ShisaD checks each action against policy, considers where the request came from, asks for confirmation when required, and records the decision. Local administrative commands use separate validation and logging rules, documented in the administrative route reference.

The core question at every action is: who asked for it? ShisaD is the user's agent — it exists to do what the user asks with the highest possible fidelity, and to prevent anything else (prompt injection, hallucination, attacker-controlled input) from taking control.

Rather than ignoring the elephant in the room, our design targets the lethal trifecta head-on: agents that access private data, process untrusted content, and take consequential actions are inherently high-risk. Most agent security research solves this by removing capabilities until the agent is safe but useless. ShisaD takes the opposite approach: keep the agent fully capable and build enforcement infrastructure that makes each capability safe to use at runtime. If a tool is insecure, the goal is to fix the enforcement, not disable the tool.

                       ┌─────────────────┐
                       │   LLM (Planner) │
   Untrusted ─────────►│   [tokens mix]  │─────────► Proposed
   Content             │                 │            Actions
                       └─────────────────┘
                                                       │
   ═══════════════════════════════════════════════════════════
   ║               ARCHITECTURAL BOUNDARY                    ║
   ═══════════════════════════════════════════════════════════
                                                       │
                                                       ▾
                        ┌─────────────────┐     ┌─────────────┐
                        │ Trusted Config  │     │  Security   │
                        │ (policies,      │     │  Analyzers  │
                        │  goals)         │     │ (metadata   │
                        └─────────────────┘     │  only)      │
                                                └──────┬──────┘
                                                       ▾
                                                APPROVE / REJECT

Features

• Long-term memory — ShisaD stores identity, current priorities, recalled facts, procedures, and supporting evidence separately in versioned local storage. Each entry records where it came from, and higher-risk additions require review
• Isolated delegated work — ShisaD keeps the main conversation active in a COMMAND session while handing delegated work—including work with untrusted, or tainted, content—to separate TASK sessions. Raw TASK output cannot enter COMMAND directly; it must pass through a summary firewall that returns a sanitized, structured summary. This separation helps prevent untrusted instructions from persisting in the long-lived COMMAND context. Each task also records its assigned scope and keeps approvals linked to the request that authorized them
• Every action is checked before execution — each action proposed by the planner or submitted through tool.execute passes through the Policy Enforcement Point (PEP), rather than being checked only when the session starts. It validates the tool and its arguments, checks permissions and resource access, restricts network destinations and credential use, scans for secrets, and applies additional rules when untrusted data reaches a sensitive destination
• Ingress and egress content controls — on ingress, ShisaD's content firewall scans untrusted input, marks it as tainted, and records where it came from. The system preserves those labels while content is processed and when it is stored. On egress, the output firewall checks content before it leaves, while policy uses the labels to determine whether related actions can proceed, require confirmation, or be blocked
• Confirmation gates, not blanket denial — user-requested actions proceed; ambiguous or tainted actions route to confirmation; only genuine anomalies trigger lockdown
• Approve actions where they were requested — on supported interfaces, users can approve or reject a pending action in the same conversation where it originated. The shisad CLI remains available as a fallback
• Detect suspicious behavior across actions — ShisaD combines signals from five independent monitors to detect patterns that a single action check might miss. It can warn the user, limit repeated requests, or enter lockdown when multiple signals indicate a genuine anomaly
• Destructive command protection — ShisaD checks shell commands before execution and blocks known destructive patterns independently of the model. In the standard supported configuration, commands run only when the required isolation is available. Operators can explicitly select expert_host_fallback to run commands without supported isolation, and ShisaD reports that weaker state
• Clean-room administrative changes — ShisaD handles changes to policy, configuration, credentials, and capabilities in a fresh SUDO session containing only the current authenticated request and trusted system state. It produces a proposal for review instead of applying changes automatically
• Messaging channels — receive and send messages through Matrix, Discord, Telegram, and Slack Socket Mode. By default, only user IDs explicitly allowed for each channel can interact with the daemon
• Built-in assistant tools — take notes, manage todos, schedule tasks, deliver results through configured channels, search and fetch web content, automate browser tasks, work with files and Git repositories, and refer to large untrusted results without loading them into the conversation history
• Keep large untrusted content out of conversation history — web pages, email bodies, and large tool results are stored separately and represented by stable references that survive daemon restarts. Raw content is loaded into an isolated, single-turn context only when needed, and terminal output is escaped to prevent control-sequence attacks
• Actions must match the user's intent — before a risky action runs, ShisaD verifies that it follows the user's intent, either directly from their request or from a plan created in a clean COMMAND session. A read involving a path that does not exist requires confirmation, while an action that would create or change an unverified missing path is blocked
• Flexible model routing — configure Shisa, OpenAI, OpenRouter, Google, and local vLLM providers with separate credentials and model choices. A single deployment can use different routes for different workloads and combine local and remote models
• Detect unexpected tool changes — reviewed local skills can add tools, and ShisaD records a fingerprint of each skill bundle and declared tool. If either changes unexpectedly, the affected skill is not loaded and the discrepancy is recorded in the audit log. Tools discovered through remote Model Context Protocol (MCP) servers are treated as untrusted and require confirmation unless the server is explicitly trusted
• Observability — comprehensive audit trail, TUI dashboard (pending actions, tasks, channel health, alerts), and doctor diagnostics

Status

ShisaD is public and under heavy development. The latest published release is v0.8.2.1.

This table tracks major release lines for reader orientation; patch releases like v0.5.1 and v0.5.2 stay in the changelog rather than being listed here.

See docs/ROADMAP.md for current direction and CHANGELOG.md for shipped release history.

Getting Started

ShisaD is under heavy development and currently intended for developers. The easiest installation path is to ask Claude Code, OpenAI Codex, or another capable coding agent to follow the deployment guide.

For complete installation and deployment instructions—including preparing the host, configuring a model provider, connecting messaging channels, and troubleshooting—see docs/DEPLOY.md. Run ShisaD on a dedicated instance or container rather than inside your development environment.

Quick Start

For a normal v0.8.2.1 installation, install the assistant package extra:

uv tool install 'shisad[assistant]'
shisad --help
shisad doctor check --component all

Run shisad with no arguments for a read-only readiness check, then use shisad setup wizard for guided interactive setup. Automated deployments can use shisad setup apply --selection FILE; it writes only when --write is provided. See docs/DEPLOY.md for the complete setup flow.

The assistant extra installs the terminal interface and client libraries for MCP, Matrix end-to-end encryption (E2EE), Discord, Telegram, and Slack. It does not enable channels or add credentials; configure those explicitly. PromptGuard remains optional, so install shisad[assistant,promptguard] only if you want its local model-based checks.

This repository also contains a tested Linux/amd64 Dockerfile candidate. No registry image is published or signed from this tree yet; build it locally and follow the volume, policy, and isolation checks in docs/DEPLOY.md.

Development checkout

Use a source checkout for development and test groups:

git clone https://github.com/shisa-ai/shisad.git
cd shisad
uv --no-config sync --frozen --group dev --group channels-runtime

YARA-backed content scanning is included in the base install through textguard[yara]. For local PromptGuard runtime checks from a source checkout, add the security runtime dependency group:

uv --no-config sync --frozen --group security-runtime --group dev --extra chat

For package installs, use the first-class PromptGuard extra, for example uv pip install 'shisad[promptguard]'. security-runtime is a uv dependency group, not a pip extra; use --group security-runtime, not --extra security-runtime. The assistant, chat, and promptguard package sets are project optional extras; assistant is the complete consumer profile, while chat remains the smaller Textual-only profile.

Memory retrieval prefers Python's sqlite3 runtime to have SQLite FTS5 enabled. ShisaD falls back when FTS5 is unavailable, but you should verify the preferred path with uv run shisad doctor check --component storage; see docs/runbooks/SQLITE.md for install guidance.

Configuration

Environment variables use SHISAD_ prefixes. Full reference: docs/ENV-VARS.md.

The v0.8.2.1 CLI can create and inspect configuration without starting the daemon:

shisad
shisad init --non-interactive
shisad config validate
shisad config upgrade
shisad config show --format human
shisad config diff --format human
shisad config schema --format json
shisad env --format human

shisad init creates one owner-only commented template at $XDG_CONFIG_HOME/shisad/config.toml (normally ~/.config/shisad/config.toml) and refuses an existing or symlink destination. It is deliberately not a setup wizard: it does not copy ambient secrets, configure a provider or policy, create daemon state, or start the daemon. Use the global --config FILE option to select another path. ShisaD resolves settings in this order, from highest to lowest priority: command-line options, environment variables, TOML configuration, and built-in defaults. Configuration display, comparison, environment, and validation output redact fields that may contain secrets.

Guided setup is available through shisad setup wizard. It displays the selected provider, policy, and channels before writing configuration. After setup, use shisad start, shisad status, and shisad tour; containers and service supervisors should use shisad start --foreground.

Chat, the one-shot terminal dashboard, and the static web snapshot share the three built-in palettes shisa-dark, shisa-light, and shisa-high-contrast. Select one with SHISAD_UI_THEME, disable optional motion with SHISAD_REDUCE_MOTION=true, and suppress palette color with NO_COLOR or a root flag such as shisad --no-color tui. Custom theme files are not supported. shisad web-ui creates a static local snapshot for investigation or export; it is not a live web application. The snapshot can contain session data, pending actions, alerts, and network-review details, so keep it private and remove it when the investigation is complete.

The shisad CLI returns exit status 0 for success, 1 for command or user-state errors, 2 for daemon connection or command-line usage errors, and 3 for invalid or unsafe configuration. shisad doctor is read-only and has no automatic repair mode. Help uses the reality-check spelling; the older realitycheck spelling remains available as a hidden compatibility alias.

Recommended: use the runner harness for local development. It handles env isolation, secret loading, and policy bootstrapping:

bash runner/harness.sh start       # background (requires tmux)
bash runner/harness.sh start --fg  # foreground
bash runner/harness.sh status
bash runner/harness.sh shisad status

The runner and plain shisad use the same per-user default socket, so another terminal can run shisad status or shisad chat without exporting SHISAD_SOCKET_PATH. See runner/README.md for details. Secrets go in runner/.env (gitignored) or SHISAD_ENV_FILE.

Manual baseline

export SHISAD_DATA_DIR="$HOME/.local/share/shisad"
case "${XDG_RUNTIME_DIR:-}" in
  /*) export SHISAD_SOCKET_PATH="$XDG_RUNTIME_DIR/shisad/control.sock" ;;
  *) export SHISAD_SOCKET_PATH="/tmp/shisad-$(id -u)/control.sock" ;;
esac
export SHISAD_POLICY_PATH="$PWD/.local/policy.yaml"
export SHISAD_LOG_LEVEL="INFO"

Provider routing

Default (Shisa.AI):

# Planner route remote-enables implicitly when SHISA key resolves.
export SHISA_API_KEY="<shisa-api-key>"

OpenAI:

export SHISAD_MODEL_REMOTE_ENABLED=true
export OPENAI_API_KEY="<openai-api-key>"
export SHISAD_MODEL_PLANNER_PROVIDER_PRESET="openai_default"
export SHISAD_MODEL_PLANNER_MODEL_ID="gpt-5.4-2026-03-05"
# Optional: export SHISAD_MODEL_PLANNER_REQUEST_PARAMETERS='{"max_completion_tokens":512}'

OpenRouter:

export SHISAD_MODEL_REMOTE_ENABLED=true
export OPENROUTER_API_KEY="<openrouter-api-key>"
export SHISAD_MODEL_PLANNER_PROVIDER_PRESET="openrouter_default"
export SHISAD_MODEL_PLANNER_MODEL_ID="qwen/qwen3.5-397b-a17b"
export SHISAD_MODEL_PLANNER_EXTRA_HEADERS='{"HTTP-Referer":"https://example.com","X-Title":"shisad"}'

Google (OpenAI-compatible):

export SHISAD_MODEL_REMOTE_ENABLED=true
export GEMINI_API_KEY="<gemini-api-key>"
export SHISAD_MODEL_PLANNER_PROVIDER_PRESET="google_openai_default"
export SHISAD_MODEL_PLANNER_MODEL_ID="gemini-3.1-pro-preview"

Local vLLM:

export SHISAD_MODEL_PLANNER_PROVIDER_PRESET="vllm_local_default"
export SHISAD_MODEL_PLANNER_BASE_URL="http://127.0.0.1:8000/v1"
export SHISAD_MODEL_PLANNER_REMOTE_ENABLED=true
export SHISAD_MODEL_PLANNER_AUTH_MODE="none"

Mixed mode (planner remote, embeddings local, monitor remote):

export SHISAD_MODEL_REMOTE_ENABLED=true

export SHISAD_MODEL_PLANNER_PROVIDER_PRESET="openrouter_default"
export SHISAD_MODEL_PLANNER_MODEL_ID="qwen/qwen3.5-397b-a17b"
export SHISAD_MODEL_PLANNER_API_KEY="<planner-openrouter-key>"

export SHISAD_MODEL_EMBEDDINGS_PROVIDER_PRESET="vllm_local_default"
export SHISAD_MODEL_EMBEDDINGS_BASE_URL="http://127.0.0.1:8000/v1"
export SHISAD_MODEL_EMBEDDINGS_REMOTE_ENABLED=true
export SHISAD_MODEL_EMBEDDINGS_AUTH_MODE="none"
export SHISAD_MODEL_EMBEDDINGS_MODEL_ID="text-embedding-3-small"

export SHISAD_MODEL_MONITOR_PROVIDER_PRESET="openai_default"
export SHISAD_MODEL_MONITOR_API_KEY="<monitor-openai-key>"
export SHISAD_MODEL_MONITOR_MODEL_ID="gpt-5.4-2026-03-05"

Verify provider setup:

uv run shisad doctor check --component provider

Auth notes:

• Use *_auth_mode=header when custom auth header names are required.
• *_auth_header_name is not accepted for *_auth_mode=bearer|none.

Channels

export SHISAD_DISCORD_ENABLED=true
export SHISAD_DISCORD_BOT_TOKEN="<token>"

export SHISAD_TELEGRAM_ENABLED=true
export SHISAD_TELEGRAM_BOT_TOKEN="<token>"

export SHISAD_SLACK_ENABLED=true
export SHISAD_SLACK_BOT_TOKEN="<xoxb-token>"
export SHISAD_SLACK_APP_TOKEN="<xapp-token>"

# Default-deny allowlist: channel -> [external_user_id]
export SHISAD_CHANNEL_IDENTITY_ALLOWLIST='{"discord":["1234567890"],"telegram":["11111"],"slack":["U12345"]}'

Assistant surfaces

# web_fetch and web_search are enabled by default.
# web_search needs a compatible JSON search backend (SearxNG-style /search?q=...&format=json).
# IP-literal, localhost, and .local/.internal/.lan backends must be in the effective web allowlist.
# Set SHISAD_WEB_ALLOWED_DOMAINS for that list; if unset, daemon policy egress hosts are used.
# Public backend hosts can be listed to preapprove backend redirects and related fetches.
# For local SearxNG setup, see docs/DEPLOY.md#web-search-backend-recommended.
export SHISAD_WEB_SEARCH_BACKEND_URL="https://search.example.com"
export SHISAD_WEB_ALLOWED_DOMAINS='["search.example.com","docs.example.com"]'

# Verify the configured tool surface from a live daemon:
# uv run python scripts/live_tool_matrix.py --tool-status

# Optional: browser automation baseline (read-mostly navigation plus
# confirmation-gated write actions). The built wheel does not install this
# source-checkout wrapper; package installs need an explicit compatible wrapper.
export SHISAD_BROWSER_ENABLED=true
export SHISAD_BROWSER_COMMAND="/path/to/shisad/scripts/shisad-playwright-cli.mjs"
export SHISAD_BROWSER_ALLOWED_DOMAINS='["example.com"]'

export SHISAD_ASSISTANT_FS_ROOTS='["/tmp/shisad-workspace"]'

Usage

Start and verify

uv run shisad start --foreground

In another shell:

uv run shisad status
uv run shisad doctor check --component all
uv run shisad doctor check --component storage
uv run shisad tui --plain

Sessions

uv run shisad session create --user alice --workspace demo
uv run shisad session list
uv run shisad session message <session-id> "summarize current priorities"

Session-derived conversation summaries can create durable memory entries by default when the session has memory.write; resulting writes are owner-scoped when the session has a complete --user / --workspace owner tuple. For clean demos or workspaces where ordinary chat should not create automatic memory, set SHISAD_MEMORY_AUTO_EXTRACTION_ENABLED=false. Operators can also raise SHISAD_MEMORY_AUTO_EXTRACTION_CONFIDENCE_THRESHOLD from 0 toward 1 to keep only higher-confidence automatic extraction proposals.

Timeline search

uv run shisad memory timeline search "what did we decide last week?" --user alice --workspace demo
uv run shisad memory timeline read <timeline-handle> --user alice --workspace demo
uv run shisad memory timeline promote <timeline-handle> --type fact --key project/decision --user alice --workspace demo

Timeline search is explicit-pull only. Results are archival evidence from prior session transcripts, not current user instructions, and shared/public channel contexts do not reveal owner-private history unless private history is explicitly allowed for that request. Read packets include role, source surface, provenance, taints, evidence refs, and content digests; search/read/promote decisions are recorded in the audit log without raw query or snippet text. For shared-channel contexts, pass the concrete room binding with --recipient, --workspace-hint, and --thread-id so same-connector/different-room history stays isolated.

Notes and todos

uv run shisad note create --key ops/runbook --content "verify doctor before deploy" --user alice --workspace demo
uv run shisad note list --user alice --workspace demo
uv run shisad todo create --title "close rollout checklist" --status open --user alice --workspace demo
uv run shisad todo list --user alice --workspace demo

Web and filesystem

uv run shisad web search "shisad security architecture" --limit 5
uv run shisad web fetch https://example.com
uv run shisad fs read /tmp/shisad-workspace/notes.txt
uv run shisad fs write /tmp/shisad-workspace/out.txt --content "hello" --confirm
uv run shisad git status --repo /tmp/shisad-workspace

Admin clean-room

uv run shisad session mode <session-id> --mode admin_cleanroom
uv run shisad channel pairing-propose --workspace <provider-workspace> --limit 50

Security Model

ShisaD assumes prompt injection will succeed and builds enforcement outside the model. On the shared planner path, the LLM is a planner, not an executor: it proposes tool calls, while runtime enforcement decides whether each call proceeds, requires confirmation, or is blocked. Those enforcement decisions do not come from the model. Authenticated local operator convenience RPCs do not claim the same PEP/control-plane path; their exact route-local boundaries are listed in docs/AUTHORITY-MAP.md.

The problem: any agent with access to private data (files, email), exposure to untrusted content (web pages, API responses), and the ability to take consequential actions (send messages, write files) is exploitable. This is the lethal trifecta. ShisaD has all three by design — it's meant to be a useful assistant, not a sandboxed demo.

The approach: instead of removing capabilities until the agent is safe (at which point you've rebuilt ChatGPT with extra steps), ShisaD keeps all capabilities available and enforces safety per-call:

• Eight checks before execution — every action from the shared planner or tool.execute is checked for tool registration, valid arguments, permissions, exposed secrets, resource access, approved destinations, credential scope, and restrictions on tainted content
• Track untrusted content — ingress controls label untrusted content as tainted and record its source. Those labels remain attached while the content is processed and stored, and egress controls use that provenance to distinguish actions arising from the user's intent, untrusted content, or unsupported model output
• Require confirmation when provenance is unclear — actions matching the user's intent can proceed when policy allows. Actions arising from ambiguous or tainted content require confirmation with context. Lockdown is reserved for behavior that the anomaly monitors identify as a threat
• Approvals stay where users can complete them — routine approvals remain in the supported interface where the request originated, using Approve/Reject controls or clear fallback instructions. When an approval requires stronger proof, ShisaD directs the user to the appropriate browser, local helper, or external signing device instead of claiming the current interface can complete it
• Behavioral anomaly detection — five independent monitors combine their signals to identify suspicious behavior that a single action check might miss
• Destructive command protection — under the supported containment profile, ShisaD checks commands against policy before execution; the model does not make this decision. A command does not run if required isolation is unavailable. Operators can explicitly choose expert_host_fallback to run without supported isolation, but that mode does not provide the same containment

Security controls preserve tool access — ShisaD's default policy keeps every built-in tool available and applies the required confirmation, isolation, credential, and audit checks when it runs. Tools that depend on external services, local software, or credentials become usable once those requirements are configured. Operators can define additional limits through a policy file selected with SHISAD_POLICY_PATH.

Egress controls — known destinations can be approved in advance. A destination explicitly requested by the user can proceed when policy allows it. A destination suggested only by untrusted content requires confirmation, while one with no connection to the user's intent is blocked.

See docs/SECURITY.md for the full security architecture and docs/DESIGN-PHILOSOPHY.md for the governing principles.

Architecture

shisad/
├── src/shisad/          # Core source
│   ├── daemon/          # Control API, handlers, runtime implementation
│   ├── security/        # PEP pipeline, content firewalls, taint tracking
│   ├── executors/       # Tool execution, egress proxy
│   ├── channels/        # Matrix, Discord, Telegram, Slack
│   ├── assistant/       # Notes, todos, web, fs/git tools
│   ├── memory/          # Structured storage with semantic search
│   ├── scheduler/       # Task scheduling and delivery
│   ├── cli/             # Click-based CLI
│   ├── ui/              # TUI dashboard
│   ├── skills/          # Hot-reloadable skill plugins
│   └── governance/      # Anomaly voting, consensus
├── tests/
│   ├── unit/            # Component tests
│   ├── integration/     # Cross-component runtime flows
│   ├── behavioral/      # Product-correctness gate
│   └── adversarial/     # Prompt injection, exfil, evasion
├── runner/              # Dev harness (tmux, env isolation, policy bootstrap)
├── scripts/             # Validation, coverage, asset checks
├── docs/                # Design docs, ADRs, runbooks, analysis
└── examples/            # Example configs and skills

Key runtime paths:

• Policy enforcement: src/shisad/security/pep.py
• Egress proxy: src/shisad/executors/proxy.py
• Handler implementation: src/shisad/daemon/handlers/_impl.py (composed from _impl_session.py, _impl_tool_execution.py, _impl_memory.py, etc.)

Development

# Example: validate one affected module and its owning tests.
uv run ruff check src/shisad/core/host_matching.py tests/unit/test_host_matching.py
uv run mypy src/shisad/core/host_matching.py
uv run --python 3.12 pytest tests/unit/test_host_matching.py -q

Ordinary changes use targeted Python 3.12 validation. Full deterministic and multi-version runs are checkpoint/release evidence, and contained suites are not rerun separately. See CONTRIBUTING.md for the validation cadence, coverage rules, and commit conventions.

Documentation

• agentic-security — literature survey on LLM agent security (78 papers, defense taxonomy, production readiness assessment)
• agentic-memory — literature survey on agent memory architectures and poisoning defenses (29+ references, attack taxonomy, defense recommendations)

License

Apache License 2.0. See LICENSE.