claw-anyllm

A plugin-shaped agent harness. Drives any LLM — including Mavis — through a Claude-CLI-style REPL. Born from a re-derivation of ultraworkers/claw-code, not a fork of its repo.

TL;DR:

• cargo build --workspace → runs claw against any LLM
• --think flag wraps the runtime in a step-by-step reasoning scaffold
• OpenHands tool delegates shell execution to a configurable HTTP backend
• Plugin manifest v2 lets you register a new LLM backend (Mavis, Ollama, anything) by writing one JSON file — no Rust change required

---

What is this?

A complete Rust workspace that implements a Claude-CLI-style agent harness where the LLM is a plugin. The conversation loop, tool execution, and permission model are all LLM-agnostic — the LLM is just another plugin entry declared in a manifest file.

You can:

1. Run it out of the box against Anthropic Claude (set ANTHROPIC_API_KEY).
2. Plug in a different LLM (Mavis, Ollama, your OpenAI-compatible API) by dropping a plugin.json manifest. No recompile.
3. Add new tools by writing a single JSON declaration. The runtime discovers and registers them.
4. Toggle "think mode" with --think for step-by-step reasoning on hard tasks.

This makes the whole thing a plugin-shaped agent harness: the host is fixed, the brain and the limbs are pluggable.

---

The plugin contract

A plugin manifest is a JSON file at .claude-plugin/plugin.json. The v2 schema supports three arrays — all optional, all backwards-compatible with v1 manifests that omit them:

{
  "schema_version": 2,
  "name": "my-plugin",
  "version": "0.1.0",
  "kind": "external",
  "providers": [
    {
      "id": "mavis",
      "display_name": "Mavis",
      "endpoint": "https://api.example.com/v1",
      "auth_env": "MAVIS_TOKEN",
      "default_model": "MiniMax-M3"
    }
  ],
  "tools": [
    {
      "name": "my_tool",
      "description": "Does something useful",
      "input_schema": { "type": "object", "properties": {} }
    }
  ],
  "hooks": [
    { "event": "PreToolUse", "command": "echo 'tool about to run'" }
  ]
}

Once a manifest is installed, the agent harness picks it up. See docs/plugin-manifest.md for the full schema and examples/ for runnable recipes.

---

Built-in plugins

Three plugins ship with the harness:

Plugin	What	How to use
Think mode	Wraps every reply in a "think first" scaffold. The runtime prepends a step-by-step reasoning directive and parses <thinking>...</thinking> blocks back out of the response.	./claw --think prompt "..."
OpenHands	A generic HTTP shell-exec backend. Sends {"code": ..., "language": ...} to a configurable endpoint. Two recipes: Bearer auth (default) or X-Session-API-Key (OpenHands Agent Server style).	OPENHANDS_ENDPOINT=... ./claw prompt "..."
Plugin manifest v2	The schema above. Already wired into the runtime's plugin loader.	Drop a manifest into .claude-plugin/

---

Build & run

# 1. Build
cd rust
cargo build --workspace

# 2. Set up auth (pick one)
export ANTHROPIC_API_KEY="sk-ant-..."          # Anthropic Claude
# or install a Mavis provider manifest:
#   cp examples/mavis-provider.json ~/.claude/plugins/mavis/plugin.json
#   export MAVIS_TOKEN="..."

# 3. Run
./target/debug/claw --help
./target/debug/claw --think prompt "say hello"
./target/debug/claw doctor

---

Try the example plugins

# Look at the Mavis example
cat examples/mavis-provider.json

# Copy to the plugin loader's search path
mkdir -p ~/.claude/plugins/mavis
cp examples/mavis-provider.json ~/.claude/plugins/mavis/plugin.json

# Set auth and run
export MAVIS_TOKEN="your-token"
./target/debug/claw --provider mavis prompt "say hello"

---

Project layout

.
├── README.md                     # this file
├── rust/                         # the canonical Rust workspace
│   ├── crates/
│   │   ├── runtime/              # conversation loop, system prompt
│   │   │   └── src/think_mode.rs        # think-mode plugin
│   │   ├── tools/                # tool registry + dispatch
│   │   │   └── src/openhands.rs         # openhands plugin
│   │   └── plugins/              # plugin manifest loader
│   │       └── src/manifest_v2.rs       # v2 manifest schema
│   └── ...
├── docs/
│   ├── plugin-manifest.md        # the v2 schema reference
│   ├── navigation-file-context.md
│   └── ...
├── examples/
│   ├── mavis-provider.json       # Mavis-style provider manifest
│   ├── ollama-provider.json      # Ollama-style provider manifest
│   └── README.md
└── LICENSE                       # MIT

---

Test status

Suite	Tests	Pass
manifest_v2	9	✓
openhands	12	✓
think_mode	9	✓
Total new	30	✓

cargo build --workspace — clean, 0 warnings.

Two pre-existing tests fail in any sandbox without PowerShell or a real MCP child process (mcp_stdio::given_child_exits_after_discovery..., tests::powershell_runs_via_stub_shell). Unrelated to this work.

---

Provenance

This repository is a re-derivation of ultraworkers/claw-code, an upstream Rust port of Claude's CLI agent harness. The intent is not to track upstream line-for-line; it's to demonstrate a plugin-shaped variant where the LLM and the tool set are first-class plugin entries.

The original code that came from upstream is retained under the same MIT license. See LICENSE and the upstream README.

---

License

MIT — same as the upstream project.

See also

• fvegiard/autoselect-skill — companion skill that auto-selects which skills to load for a given turn.
• fvegiard/claw-code — the upstream-derivation fork where this work was first done.
• ultraworkers/claw-code — the original.