loop.go: - NewAgentLoop now takes a context.Context and returns (*AgentLoop, error); memory init is mandatory — errors are returned instead of logged+skipped - Wire IdentitySync into the loop (identitySync field) - Use cfg.SandboxPath() / cfg.DBPath() / cfg.RestrictToSandbox() - SetContextWindow() passed to ContextBuilder from cfg.Agents.Defaults.MaxTokens context.go: - NewContextBuilder resolves skills dirs via config.SkillsDir() and config.ConfigDir() (XDG-aware) instead of hardcoded ~/.picoclaw - BuildSystemPrompt: priority-ordered sections with token-budget trimming (charsPerToken heuristic); skills section updated with tool_search guidance pkg/agent/toolloop.go (new): - Extract ToolLoopMode, DAGRunResult, DAGRunFunc, RouteFunc types from pkg/tools/toolloop.go to break the tools→dag→securebus→tools import cycle |
||
|---|---|---|
| .github | ||
| assets | ||
| cmd/picoclaw | ||
| config | ||
| docs | ||
| eval | ||
| internal/fantasy | ||
| pkg | ||
| scripts | ||
| workspace | ||
| .dockerignore | ||
| .env.example | ||
| .gitignore | ||
| .goreleaser.yaml | ||
| docker-compose.yml | ||
| Dockerfile | ||
| Dockerfile.goreleaser | ||
| go.mod | ||
| go.sum | ||
| LICENSE | ||
| Makefile | ||
| README.ja.md | ||
| README.md | ||
| README.pt-br.md | ||
| README.zh.md | ||
| ROADMAP.md | ||
PicoClaw
A managed fork of sipeed/picoclaw — an ultra-lightweight AI agent runtime written in Go.
This fork diverges from upstream with its own architectural decisions:
- vendored LLM SDK
- MemGPT-style tiered memory
- Isolated Tool Runtime:
- capability-based security
- DAG-based parallel tool execution
- libSQL-native storage with vector search
Note
Where it makes sense, upstream changes are merged back. This is not a strict rule, but a guideline.
Why This Fork Exists
The upstream PicoClaw project is a solid foundation — a single-binary AI agent that runs on $10 hardware with <10MB RAM. But it has architectural gaps that limit extensibility:
- No privilege boundary between the LLM and tool execution — a compromised tool has full process access
- No structured memory beyond flat markdown files
- All tools loaded into context every request (token waste at scale)
- Sequential tool calling only — each tool call requires a full inference pass
- Hand-rolled LLM provider implementations with no streaming, retry, or multi-provider support
This fork addresses all of those while preserving the original's strengths: small binary, low memory, single-process deployment.
Architecture
flowchart TB
subgraph AgentLoop["Agent Loop"]
AL[AgentLoop] --> CB[ContextBuilder]
AL --> TR[ToolRegistry]
AL --> MS[MemoryStore]
AL --> FSM[ReAct FSM]
end
subgraph ITR["Isolated Tool Runtime"]
SB[SecureBus] --> CAP[Capability Check]
SB --> SI[Secret Injection]
SB --> EX[Tool Execution]
SB --> LS[Leak Scanning]
SB --> AU[Audit Log]
end
subgraph DAG["DAG Executor"]
DE[Executor] --> RES[Dependency Resolver]
DE --> WAVE[Parallel Wave Dispatch]
DE --> JOIN[Joiner Synthesis]
RT[Router] -->|ModeReAct| FSM
RT -->|ModeDAG| DE
end
subgraph Fantasy["Fantasy SDK (vendored)"]
FA[FantasyAdapter] --> FP[Provider Registry]
FP --> OR[OpenRouter / Anthropic / OpenAI / Gemini]
end
subgraph Progressive["Progressive Disclosure"]
TR --> TS["tool_search (fuzzy)"]
TR --> TC["tool_call (dispatch)"]
end
subgraph Memory["Memory System"]
MS --> WC[Working Context]
MS --> RC[Recall Memory]
MS --> AR[Archival Memory]
MS --> OBS[Observational Memory]
MS --> DAGC[DAG Compression]
end
subgraph Storage["libSQL Storage"]
DEL[LibSQLDelegate] --> BLOB["BLOB PK (UUIDv7)"]
DEL --> F32["F32_BLOB (embeddings)"]
DEL --> FTS["FTS5 + BM25"]
DEL --> VEC["vector_top_k (ANN)"]
end
subgraph Security["Security"]
VLT[Vault XChaCha20] --> SS[SecretStore]
SS --> KR[Keyring / Env / File]
RED[Redactor] --> SB
ZKP[Schnorr ZKP] -.-> SOCK[Daemon Socket]
end
subgraph Bus["Message Bus"]
BUS[MessageBus] --> TG[Telegram]
BUS --> DC[Discord]
BUS --> SL[Slack]
BUS --> LN[LINE]
BUS --> DT[DingTalk]
BUS --> QQ[QQ]
end
AL --> SB
AL --> FA
MS --> DEL
AL --> BUS
SB --> TR
style AgentLoop fill:#1a1a2e,stroke:#e066ff,stroke-width:2px,color:#fff
style ITR fill:#1a1a2e,stroke:#ff6b6b,stroke-width:2px,color:#fff
style DAG fill:#1a1a2e,stroke:#ffab00,stroke-width:2px,color:#fff
style Fantasy fill:#1a1a2e,stroke:#4d94ff,stroke-width:2px,color:#fff
style Progressive fill:#1a1a2e,stroke:#00bfa5,stroke-width:2px,color:#fff
style Memory fill:#1a1a2e,stroke:#2eb82e,stroke-width:2px,color:#fff
style Storage fill:#1a1a2e,stroke:#9c27b0,stroke-width:2px,color:#fff
style Security fill:#1a1a2e,stroke:#ff9800,stroke-width:2px,color:#fff
style Bus fill:#1a1a2e,stroke:#607d8b,stroke-width:2px,color:#fff
Key Design Decisions
| Decision | Rationale |
|---|---|
| Isolated Tool Runtime | All tool calls route through a SecureBus that enforces capability manifests, injects secrets, scans output for leaks, and writes audit logs. The LLM never sees raw secrets. See ADR-001. |
| DAG executor | LLMCompiler-style parallel tool dispatch. The planner builds a dependency DAG in a single inference pass; the executor dispatches independent nodes concurrently. Joiner synthesizes results. Replanning on failure. Falls back to ReAct for simple single-tool cases. |
| Vendored Fantasy SDK | charm.land/fantasy vendored into internal/fantasy/ via go.mod replace directive. Enables direct modification for streaming hooks, tool call repair, and progressive disclosure. |
| MemGPT + Observational Memory | Working context (hot), recall items (warm), archival chunks (cold, embedded + indexed), plus observational memory (compressed conversation history with priority-tagged observations and temporal reasoning). DAG-based context budget compression manages token allocation across tiers. |
| Progressive tool disclosure | Agent sees only tool_search and tool_call meta-tools. Discovers actual tools on demand via fuzzy search. Cuts system prompt tokens for large registries. |
| libSQL over modernc/sqlite | Native F32_BLOB for vector storage, libsql_vector_idx for ANN search, FTS5 for full-text. Single database, no external vector DB dependency. |
| BLOB primary keys | 16-byte UUIDv7 stored as BLOB. Compact, byte-comparable, monotonically sortable by creation time. |
| XChaCha20-Poly1305 vault | Secrets encrypted at rest with AES-256-GCM or XChaCha20-Poly1305. Master key from OS keyring, env var, or file. Schnorr ZKP for daemon-mode authentication. |
| Goose migrations | Schema managed by pressly/goose/v3. 10 versioned migrations covering core schema, FTS5, vector indexes, KV store, documents, audit log, conversations, runtime state, jobs, and conversation graphs. |
| FlatBuffers command protocol | Zero-copy serialized ToolRequest/ToolResponse for the ITR command vocabulary. Same binary format across in-process channels, Unix sockets (daemon mode), and wazero WASM host calls. |
Project Layout
cmd/picoclaw/ # CLI entrypoint
internal/fantasy/ # Vendored charm.land/fantasy SDK
docs/adr/ # Architecture Decision Records
eval/ # Promptfoo-based evaluation harness
pkg/
├── agent/ # Agent loop, ReAct FSM, context builder
│ ├── conversations/ # Conversation store (multi-turn tracking)
│ ├── mentions/ # Mention tracking
│ └── threads/ # Thread store
├── auth/ # OAuth2 + PKCE for provider auth
├── bus/ # Hub-and-spoke message bus
├── cache/ # Generic LRU+TTL cache (SWR, tag invalidation)
├── channels/ # Telegram, Discord, Slack, LINE, DingTalk, QQ, MaixCAM
├── config/ # JSON config with env var overrides, XDG paths
├── cron/ # Cron scheduler (gronx-based)
├── devices/ # Hardware device hotplug (USB on Linux)
├── fantasy/ # Fantasy SDK adapter (provider factory, type conversion)
├── health/ # HTTP health/readiness endpoints
├── heartbeat/ # Periodic task execution
├── ids/ # UUIDv7 generation + BLOB codec
├── itr/ # Isolated Tool Runtime
│ ├── dag/ # DAG executor, planner, resolver, router, replanner
│ ├── itrfb/ # FlatBuffers generated code (command protocol)
│ └── wasm/ # wazero WASM isolate runtime + transport
├── logger/ # Structured logger
├── memory/ # Memory system
│ ├── dag/ # DAG-based context budget compression
│ ├── delegate/ # libSQL storage backend (FTS5, vector, capabilities)
│ ├── migrations/ # Goose versioned schema migrations (001–010)
│ ├── observation/ # Observational memory (observer, reflector, store)
│ ├── sqlc/ # sqlc config + generated code
│ └── store/ # MemoryStore, retrieval, chunking, scoring, queuing
├── messages/ # Canonical message/tool-call types
├── pcerrors/ # Structured error types
├── rlm/ # Recursive Language Model engine (rope, fanout, strategy)
├── security/ # Vault, SecretStore, Redactor, URL guard, Schnorr ZKP
│ └── securebus/ # SecureBus (policy, audit, transport, socket transport)
├── session/ # Session manager with LRU disk-backed cache
├── skills/ # Skill loader, installer, dependency graph, templates
├── tools/ # Tool registry, meta-tools, built-in tools, CapableTool
├── voice/ # Groq Whisper voice transcription
└── worker/ # Background job worker
skills/ # Built-in skills (weather, tmux, summarize, github, hardware)
config/ # Example configuration files
Quick Start
Build from source
Important
Requires
CGO_ENABLED=1— the go-libsql driver links against glibc.
git clone https://github.com/ZanzyTHEbar/picoclaw.git
cd picoclaw
make build
Configure
./bin/picoclaw onboard
The onboard wizard initializes config, workspace, and optionally sets up encrypted secret storage.
Edit ~/.picoclaw/config.json:
{
"agents": {
"defaults": {
"model": "anthropic/claude-sonnet-4-20250514",
"max_tokens": 8192,
"temperature": 0.7,
"max_tool_iterations": 20
}
},
"providers": {
"openrouter": {
"api_key": "sk-or-v1-xxx",
"api_base": "https://openrouter.ai/api/v1"
}
},
"tools": {
"progressive_disclosure": true,
"web": {
"duckduckgo": { "enabled": true, "max_results": 5 }
}
}
}
Run
# One-shot
picoclaw agent -m "What is 2+2?"
# Interactive REPL
picoclaw agent
# Gateway (Telegram, Discord, etc.)
picoclaw gateway
Docker
cp config/config.example.json config/config.json
docker compose --profile gateway up -d
docker compose logs -f picoclaw-gateway
Secret Management
PicoClaw encrypts secrets at rest with XChaCha20-Poly1305. The master key is sourced from an environment variable, OS keyring, or file.
picoclaw secret init # Generate a master key
picoclaw secret add <name> # Store a secret (interactive prompt)
picoclaw secret list # List secret names
picoclaw secret delete <name> # Remove a secret
Warning
This is a security-sensitive operation. The master key is used to encrypt and decrypt secrets. If you lose it, you will not be able to decrypt secrets. You should should NEVER store the master key in a file or environment variable if possible.
Set the master key: export PICOCLAW_MASTER_KEY=<hex>
Tools declare which secrets they need via CapableTool.Capabilities(). The SecureBus injects secrets into tool execution context at runtime — the LLM never sees them. Tool output is scanned for leaked patterns before it reaches the agent loop.
Daemon Mode
For non-embedded deployments, the SecureBus can run in a separate privileged daemon process. The agent connects as an unprivileged client over a Unix domain socket.
picoclaw daemon start # Start daemon (foreground, Ctrl+C to stop)
picoclaw daemon status # Check if running
picoclaw daemon stop # Stop a running daemon
LLM Providers
The Fantasy SDK handles provider routing. Configure any supported provider:
| Provider | Config Key | Notes |
|---|---|---|
| OpenRouter | openrouter |
Access to all models via single API key |
| Anthropic | anthropic |
Claude direct |
| OpenAI | openai |
GPT direct |
| Google Gemini | gemini |
Gemini direct |
| Groq | groq |
Fast inference + Whisper voice transcription |
API key links: OpenRouter · Anthropic · OpenAI · Gemini · Groq
Chat Channels
| Channel | Setup Complexity |
|---|---|
| Telegram | Easy — single bot token |
| Discord | Easy — bot token + message content intent |
| Easy — AppID + AppSecret | |
| DingTalk | Medium — app credentials |
| LINE | Medium — credentials + webhook URL |
| Slack | Medium — app credentials + event subscriptions |
Telegram
- Create bot via
@BotFatheron Telegram, copy token - Get your user ID from
@userinfobot - Configure:
{
"channels": {
"telegram": {
"enabled": true,
"token": "YOUR_BOT_TOKEN",
"allowFrom": ["YOUR_USER_ID"]
}
}
}
- Run
picoclaw gateway
Discord
- Create application at https://discord.com/developers/applications
- Create bot, copy token, enable MESSAGE CONTENT INTENT
- Get your User ID (Developer Mode → right-click avatar → Copy User ID)
- Configure:
{
"channels": {
"discord": {
"enabled": true,
"token": "YOUR_BOT_TOKEN",
"allowFrom": ["YOUR_USER_ID"]
}
}
}
- Invite bot: OAuth2 → URL Generator → Scopes:
bot→ Permissions:Send Messages,Read Message History - Run
picoclaw gateway
- Create application at QQ Open Platform
- Configure:
{
"channels": {
"qq": {
"enabled": true,
"app_id": "YOUR_APP_ID",
"app_secret": "YOUR_APP_SECRET",
"allow_from": []
}
}
}
- Run
picoclaw gateway
DingTalk
- Create internal app at DingTalk Open Platform
- Configure:
{
"channels": {
"dingtalk": {
"enabled": true,
"client_id": "YOUR_CLIENT_ID",
"client_secret": "YOUR_CLIENT_SECRET",
"allow_from": []
}
}
}
- Run
picoclaw gateway
LINE
- Create Messaging API channel at LINE Developers Console
- Configure:
{
"channels": {
"line": {
"enabled": true,
"channel_secret": "YOUR_CHANNEL_SECRET",
"channel_access_token": "YOUR_CHANNEL_ACCESS_TOKEN",
"webhook_host": "0.0.0.0",
"webhook_port": 18791,
"webhook_path": "/webhook/line",
"allow_from": []
}
}
}
- Set up HTTPS webhook (e.g.,
ngrok http 18791) and configure the URL in LINE console - Run
picoclaw gateway
Memory System
PicoClaw implements a multi-tier memory system combining MemGPT-style tiered storage with observational memory compression:
| Tier | Purpose | Storage | Search |
|---|---|---|---|
| Working Context | Current focus, active goals | Single JSON document per agent | Direct load |
| Recall Memory | Session-scoped conversation items | Rows with metadata + timestamps | FTS5 + BM25 |
| Archival Memory | Long-term knowledge, chunked + embedded | F32_BLOB embeddings + FTS5 index | Vector ANN + FTS5 fusion (RRF) |
| Observational Memory | Compressed conversation history | Priority-tagged observations with 3-date model | Prefix-cacheable block |
| DAG Compression | Hierarchical context summaries | Tree nodes with lossless pointers to originals | Budget-allocated traversal |
The agent interacts with memory through a unified memory tool. Large tool results are automatically offloaded to archival memory. Retrieval uses Reciprocal Rank Fusion (RRF) to combine vector similarity and full-text relevance, with recency decay and metadata pre-filtering.
Schema is managed by Goose with 10 versioned migrations.
CLI Reference
| Command | Description |
|---|---|
picoclaw onboard |
Initialize config, workspace, and secret storage |
picoclaw agent -m "..." |
One-shot chat |
picoclaw agent |
Interactive REPL |
picoclaw gateway |
Start message bus gateway |
picoclaw status |
Show system status (incl. memory) |
picoclaw memory |
Memory system management |
picoclaw secret <sub> |
Secret management (init, add, list, delete) |
picoclaw daemon <sub> |
Daemon management (start, stop, status) |
picoclaw cron list |
List scheduled jobs |
picoclaw cron add ... |
Add a scheduled job |
picoclaw skills <sub> |
Skill management (install, list, remove) |
Security
Workspace Sandbox
File and command access is restricted to the configured workspace by default. The exec tool blocks dangerous commands regardless of sandbox setting.
{ "agents": { "defaults": { "restrict_to_workspace": false } } }
Isolated Tool Runtime (ITR)
The SecureBus mediates all tool execution. Tools declare capabilities via the CapableTool interface — secrets needed, network endpoints, filesystem paths, shell access level. Tools that don't implement it get zero capabilities.
The pipeline: capability check → secret injection → tool execution → leak scanning → audit log.
See ADR-001 for the full design including DAG executor convergence, RLM integration, and the FlatBuffers command protocol.
Development
make build # Build for current platform (output: bin/)
make build-all # Cross-compile (linux/amd64, linux/arm64, linux/riscv64, darwin/arm64, windows/amd64)
make install # Install to ~/.local/bin + copy skills
make fmt # go fmt ./...
make deps # go get -u + go mod tidy
make clean # Remove build artifacts
sqlc
Memory queries are generated by sqlc. After modifying SQL files:
cd pkg/memory/sqlc && sqlc generate
CI enforces that generated code matches: sqlc generate + git diff --exit-code.
Migrations
Schema changes go through Goose migrations in pkg/memory/migrations/. Migrations run automatically on startup.
Evaluation
A promptfoo-based evaluation harness lives in eval/:
cd eval && go run ./cmd/eval-runner
Syncing Upstream
git fetch upstream
git merge upstream/main
See internal/fantasy/VENDORING.md for syncing the vendored Fantasy SDK.
Roadmap
See ROADMAP.md for the full project roadmap covering context management, skill graphs, agentic retrieval, delegation protocol, and multi-agent coordination.
Upstream
This is a fork of sipeed/picoclaw, originally inspired by nanobot. The upstream project targets $10 RISC-V hardware with <10MB RAM — a constraint this fork respects while extending the agent's cognitive and security architecture.
License
MIT — see LICENSE.