BendClaw
Distributed AI teams. Share everything. Co-evolve.
A Rust-native runtime where every execution produces reusable knowledge — no prompt engineering required.
The engine behind evot.ai
Most agent frameworks are stateless — every run starts from zero. BendClaw turns each execution into reusable knowledge that's automatically shared across the team and injected into future runs. Agents co-evolve without prompt engineering.
It's also a distributed runtime from day one. Agents collaborate across nodes, fan out subtasks, and collect results — all on a single shared data layer backed by Databend. Add nodes to scale; the cluster handles coordination.
- Autonomous learning — agents extract and share knowledge from every run, auto-injected into future prompts for continuous co-evolution
- Cluster dispatch — agents collaborate across nodes, fan out subtasks, and collect results
- Lease-based scheduling — tasks and channel receivers claimed via distributed DB leases; automatic failover
- Shared persistent memory — vector + full-text search on shared cloud storage, one unified data layer
- Hub integrations — 100+ integrations (GitHub, Slack, Email, etc.) via pluggable skills
- Secret-safe execution — secrets in a vault, never exposed to LLMs; injected only at tool execution time
- Full traceability — spans, events, audits, sensitive field redaction; humans review, agents execute
- Multi-tenant isolation — separate DB per agent, isolated workspace per session
- 32 built-in tools — file, search, shell, memory, recall, task, web, databend, channel, cluster
- 50+ REST endpoints — SSE streaming, Bearer auth, per-agent scoping
- Go to app.evot.ai → Add AgentOS → Your Computer
 |
 |
- Copy and run the setup command:
curl -fsSL https://app.evot.ai/api/setup | sh -s -- <BASE64_CONFIG>bendclaw run— the console detects your instance automatically
CLI commands:
| Command | Description |
|---|---|
bendclaw run |
Run in foreground (default) |
bendclaw start |
Start server in background |
bendclaw stop |
Stop the server |
bendclaw restart |
Kill old process, start new one |
bendclaw status |
Show server status |
bendclaw update |
Upgrade to latest release |
For development from source, see Development.
HTTP / SSE HTTP / SSE HTTP / SSE
│ │ │
▼ ▼ ▼
┌───────────────────┐ ┌───────────────────┐ ┌───────────────────┐
│ BendClaw Node A │ │ BendClaw Node B │ │ BendClaw Node N │
│ │ │ │ │ │
│ Gateway │ │ Gateway │ │ Gateway │
│ Kernel + Hub │ cluster │ Kernel + Hub │ cluster │ Kernel + Hub │
│ Recall │◄───────▶│ Recall │◄───────▶│ Recall │
│ Lease │ RPC │ Lease │ RPC │ Lease │
│ Channels │ │ Channels │ │ Channels │
│ │ │ │ │ │
└─────────┬─────────┘ └─────────┬─────────┘ └─────────┬─────────┘
└─────────────────────────────┼─────────────────────────────┘
▼
┌───────────────────────────────────────────────────────┐
│ Databend (Cloud) │
│ │
│ sessions · runs · memories (vector + FTS) │
│ learnings · knowledge · skills · traces │
│ tasks · config · variables · feedback · channels │
│ │
│ Shared cloud storage. All agents, │
│ one data layer. │
└───────────────────────────────────────────────────────┘
| Layer | Role |
|---|---|
| Gateway | HTTP routing, SSE streaming, Bearer auth, CORS, request logging |
| Kernel | Agent loop, LLM router (Anthropic / OpenAI) with circuit breaker and failover, tool registry, context compaction, prompt builder |
| Recall | Post-run knowledge extraction (fire-and-forget), learning accumulation, auto-injection into future prompts |
| Lease | Distributed lease coordination — claim/renew/release across nodes; per-resource callbacks for task scheduling and channel receiver lifecycle |
| Hub | Pluggable skill registry, auto-sync from remote repo, 100+ integrations fed into Kernel |
| Channels | Webhook ingestion (Feishu, Telegram, GitHub), HTTP API channel, lease-managed receivers, centralized sender trust (allow_from), inbound dispatch to Kernel |
| Cluster | Peer-to-peer RPC — node registration, heartbeat, autonomous subtask dispatch across nodes |
| Databend | Shared cloud storage — all agent data, one unified data layer |
| Category | Tools | Description |
|---|---|---|
| File | file_read, file_write, file_edit, list_dir |
Workspace file operations (sandbox mode optional) |
| Search | grep, glob |
Content search (regex) and file search (glob pattern) across workspace |
| Shell | shell |
Allowlisted commands with configurable timeout |
| Memory | memory_write, memory_read, memory_search, memory_list, memory_delete |
Long-term memory with vector + full-text search |
| Skill | skill_read, create_skill, remove_skill |
Skill documentation access and management; executable skills can act as runtime tools |
| Recall | learning_write, learning_search, knowledge_search |
Agent self-improvement: write learnings, search accumulated knowledge |
| Task | task_create, task_list, task_get, task_update, task_delete, task_toggle, task_history |
Cron task self-management |
| Web | web_search, web_fetch |
Web search and page fetching |
| Databend | databend |
SQL queries against the agent's Databend database |
| Channel | channel_send |
Send messages through connected channels |
| Cluster | cluster_nodes, cluster_dispatch, cluster_collect |
Discover peers, dispatch subtasks to other agents, collect results |
| Coding | claude_code, codex_exec |
Delegate coding tasks to Claude Code or OpenAI Codex CLI with multi-turn session support |
50+ REST endpoints with SSE streaming, Bearer auth, and per-agent scoping. See docs/API.md for full reference.
make setup # install Rust toolchain, git hooks
make check # fmt + clippy
make test # unit + integration + contract (no credentials needed)
make coverage # generate HTML coverage report- GitHub Issues — bug reports & feature requests
- Twitter @Evot_AI — updates & announcements
- team@evot.ai — reach the team directly
Apache-2.0