FireClaw

FireClaw is a terminal-based AI coding assistant and agent framework built around local execution. Its public repository describes interactive and single-shot prompting, file and code editing, shell execution, code search, web fetch, session persistence, Git awareness, project instructions through CLAW.md and cost tracking.¹

Extensibility is a major part of the design. FireClaw documents MCP connections over stdio, WebSocket and HTTP, including lifecycle management and OAuth proxy support, alongside a plugin system, pre/post tool hooks and specialized skills. The permission model separates read-only, workspace-write and full-access modes and supports per-tool denylists.¹

The project presents three implementations: Mojo as the primary direction, a Rust port and a Python reference implementation. The current architecture uses Python interoperability for networking, TLS, WebSocket and OAuth in the Mojo path. Anthropic is the primary model-provider target; broader provider support including OpenAI, Gemini and local models is listed on the roadmap rather than as a completed current capability.¹

The headline figures of 184+ tools, 207 commands, 274 Rust tests and near-C/native-speed advantages are reported by the project itself. They describe the intended and implemented surface but should not be read as independent comparative benchmarks.¹

FireClaw is free, self-hosted open-source software under Apache-2.0. There is no FireClaw subscription required for the repository itself. Users provide their own model-provider credentials, so actual operating cost depends mainly on the selected external LLM and any external MCP services.¹

The documented quick starts require local developer tooling. The primary Mojo route uses the Max toolchain and Pixi; the Rust port requires Rust/Cargo; the Python implementation uses a local Python environment. This is substantially more technical than installing a mainstream IDE extension, but it gives users direct control over code, runtime and permissions.¹

FireClaw's transparency is helped by its public source, Apache-2.0 license and explicit permission architecture. The repository exposes the implementation structure and documents read-only, workspace-write and full-access modes, denylists and tool hooks. Those controls matter because the agent can edit files and execute shell commands locally.¹

Maturity is the main evidence limitation. The public repository currently shows only a very short commit history, while the same README's roadmap says Mojo-native async and HTTP/TLS remain future work and identifies remaining Mojo tool-parity gaps. The project also makes strong speed, production-readiness and competitive tool-count claims without independent benchmarks. AiToolMap therefore treats those as author claims rather than verified performance facts.¹

The external source audit found no current exact-product independent benchmark, substantive review or eligible user-rating corpus for Eldergenix FireClaw. Searches also surfaced unrelated game and fiction uses of the name “Fireclaw,” which were rejected as product mismatches. Missing external coverage is neutral rather than a negative performance signal.

FireClaw is most relevant to developers and agent-framework experimenters who want a local terminal agent they can inspect, modify and extend, especially users interested in Mojo, MCP and permissioned local automation. It also fits developers who prefer a BYOK model and want the option to work across Mojo, Rust or Python implementations.¹²

It is a weaker fit for teams that want a mature supported commercial product, a simple graphical installation, broadly completed multi-model support or independently demonstrated performance and reliability. Production adopters should evaluate the exact runtime they intend to use rather than assuming feature parity from the project's aggregate feature tables.