clawboo-native) is the fastest way to get a working team of agents. It is built directly into the Clawboo server — no CLI to install, no Gateway to run. Paste a provider key (or point it at a local Ollama instance), and your agents are ready to execute tasks. It supports Anthropic, OpenAI, OpenRouter, and Ollama, and it keeps a persistent conversation memory per agent across every run.
What it gives you
- Zero install.
clawboo-nativeships inside the Clawboo server. There is nothing to install beyond Clawboo itself. - Multi-provider. Connect an Anthropic key, an OpenAI key, an OpenRouter key, a local Ollama instance, or any combination. Each agent can have its own primary provider with fallbacks.
- Persistent memory. Each agent’s conversation transcripts live in a stable per-identity home on disk. A resumed run picks up exactly where the prior session ended.
- Full shared-plane access. Native agents attach Clawboo’s Tasks, Memory, Tools, and TeamChat MCP servers in-process — the same coordination surface every other runtime uses.
- Real per-turn cost tracking. The native runtime emits cost events after each model turn, not just once at the end, so the budget kill-switch sees live spend mid-run.
Capabilities
| Capability | Value |
|---|---|
| Streaming (live text deltas) | ✅ |
| MCP (Clawboo’s tool servers) | ✅ |
| Worktrees (isolated git workspace) | ✅ |
| Resume (continue a prior session) | ✅ |
| Tool approval gates | ✅ |
| Context window | 200,000 tokens |
| Persistent per-identity home | ✅ |
Supported providers
| Provider | Credential | Notes |
|---|---|---|
| Anthropic | ANTHROPIC_API_KEY | Default slot when no provider is specified |
| OpenAI | OPENAI_API_KEY | Uses OpenAI Chat Completions with streaming |
| OpenRouter | OPENROUTER_API_KEY | Same OpenAI-compatible client with a base-URL override |
| Ollama | (none — keyless) | Reaches your local Ollama at http://localhost:11434/v1 by default; set OLLAMA_BASE_URL to override |
primaryProvider, with declared fallbacks tried in order when a turn fails before any output is emitted.
Clawboo Native is a peer runtime, not a universal fallback. It does not convert, migrate, or export agents from other runtimes. A team can freely mix Native agents with agents on Claude Code, Codex, Hermes, or OpenClaw.
Connect from the Runtimes panel
Open the Runtimes panel
In the Clawboo dashboard, navigate to Runtimes. The Clawboo Native card is always present — it never shows
not-installed because there is nothing to install.(Optional) Verify your key before saving
Click Verify key on the card and paste your API key. Clawboo makes a single lightweight request to the provider’s models endpoint and tells you whether the key is valid — without storing it. This step is optional but useful for catching typos.
Paste your provider key
Click Connect and paste your API key. Select the provider that matches your key from the dropdown (Anthropic is the default). Clawboo stores the key in the encrypted vault and the card flips to Ready.For Ollama, no key is required — just ensure Ollama is running locally. The
provider: "ollama" connect call is a no-op on storage.Seed a starter team (first run)
If this is your first time setting up Clawboo Native, use the onboarding wizard or call the seed endpoint directly. Clawboo creates a leader agent and a specialist agent pre-configured for your chosen provider.Per-provider default models: Anthropic →
claude-sonnet-4-6 / claude-haiku-4-5; OpenAI → gpt-4o / gpt-4o-mini; OpenRouter → anthropic/claude-haiku-4.5 / openai/gpt-4o-mini; Ollama → llama3.2 / llama3.2.Persistent conversation memory
Each native agent gets a stable home at<clawboo home>/runtimes/clawboo-native/<agentId>/. After every run, Clawboo saves the full conversation transcript to that home. When you resume the same agent on the same runtime, the prior transcript is reloaded and the conversation continues — not just a prose summary, but the actual session history.
This persistence is what resume: true means for the native runtime. Cross-runtime continuity (picking up a task on a different runtime) uses the prose handoff note in AGENT_HANDOFF.json instead.
Clawboo’s verification critic intentionally runs without a home — the reviewer never shares the builder’s persisted transcripts, so builder and judge stay independent.
Built-in file tools
When a native agent works on a coding task in a worktree, it gets three file tools automatically:| Tool | What it does |
|---|---|
read_file | Reads a UTF-8 file relative to the workspace root (capped at 64 KiB) |
write_file | Writes a UTF-8 file, creating parent directories as needed |
list_files | Lists entries in a directory (defaults to the workspace root) |
.. or an absolute path is rejected. Agents working on research or review tasks (no worktree) receive no file tools.
Provider fallback
If a provider call fails before any output is emitted, Clawboo automatically tries the next candidate in the agent’sfallbacks list. Once a candidate starts streaming, that provider is locked in for the rest of the run — mid-stream retries are not possible without duplicating output. Providers with no resolvable key are skipped automatically.
Fallback-worthy errors: authentication failure, rate limit, overload, and network errors. A mid-stream error surfaces as-is.
Verify it worked
After connecting,GET /api/runtimes should show clawboo-native with connectionState: "ready" and health.ok: true. Native health is true when any provider key resolves (ANTHROPIC_API_KEY, OPENAI_API_KEY, OPENROUTER_API_KEY) or OLLAMA_BASE_URL is set — no binary probe, no network call at status time.
Troubleshooting
See also
- Runtimes overview — the full capability matrix across all five runtimes
- Connecting runtimes — the install/connect/disconnect lifecycle and the encrypted vault
- Claude Code · Codex · Hermes · OpenClaw — the other runtimes