Crabtalk
A daemon for running agents, dispatching tools, and connecting MCP servers.
- Source repo
- crabtalk/crabtalk-20261010
- Stars
- ★ 737
- Last updated
- 2mo ago
- License
- Apache-2.0
- Primary language
- Rust
- FA score
- 27/100 · Major gaps
At a glance
- How it runs
- Works with
- Portable with changes
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- An engineer who needs a persistent local agent service and wants to converse with it through a terminal UI.
- Not a fit if
- Users unwilling to install the Rust/Cargo toolchain
- Production teams needing documented OS support and recovery behavior
- Source review
- 27/100 · Major gaps 4 safety controls not found
What does this agent do, and when should you use it?
Crabtalk is an agent runtime delivered as a daemon that runs agents, dispatches tool calls, and connects to MCP servers. Its built-in tools are shell, task delegation, and memory, and it also supports skills supplied as Markdown prompt files. Components are obtained with crabup, with crabtalkd as the daemon and crabtalk-tui as the interactive client. Apps can be standalone binaries that connect to the daemon through auto-discovery. Adopters must complete interactive LLM endpoint setup, while the supplied material does not name supported model providers or authentication methods.
Install crabup, pull the daemon and TUI packages, then run crabtalkd setup to configure an LLM endpoint. Once started, the daemon runs agents and dispatches its built-in shell, task delegation, and memory tools, or connects to MCP servers. Users invoke crabtalk-tui to chat with the daemon. Apps under apps/ are standalone binaries that connect through auto-discovery, while skills are supplied as Markdown prompt files.
- An engineer who needs a persistent local agent service and wants to converse with it through a terminal UI.
- A development team that wants agents to work with capabilities exposed by existing MCP servers alongside built-in tools.
- A tool builder who wants one daemon to coordinate shell execution, task delegation, and memory.
- A systems or Rust-oriented team building standalone binaries that connect to a shared agent daemon through auto-discovery.
- A developer who wants to maintain reusable skills as Markdown prompt files.
How do you install or deploy this agent?
Install the package manager: cargo install crabup. Fetch the components: crabup pull daemon && crabup pull tui. Run crabtalkd setup and complete its interactive LLM endpoint configuration. Install and start the service with crabup daemon start. The first working client invocation is crabtalk-tui. The supplied material does not specify endpoint credentials, provider choices, or configuration fields.
How do you use this agent?
After completing crabtalkd setup, run crabup daemon start to install the service unit and start the daemon, then run crabtalk-tui to chat. The repository identifies crates/crabtalk/config.toml as the full configuration reference. Extend the deployment through MCP server integration, skills represented by Markdown prompt files, or standalone apps described under apps/.
What are this agent's strengths and limitations?
- Combines agent execution, tool dispatch, and MCP server connections in one daemon-oriented runtime.
- Explicitly includes shell, task delegation, and memory as built-in tools.
- Provides both a crabtalk-tui interactive client and an auto-discovery boundary for standalone apps.
- Uses crabup to fetch the daemon and TUI as separate packages.
- LLM endpoint setup is required through crabtalkd setup, but supported providers, models, and authentication are not identified.
- The documented installation path depends on Cargo and crabup, adding adoption work outside Rust/Cargo environments.
- The supplied material does not document supported operating systems, resource requirements, service recovery, or troubleshooting behavior.
- MCP integration is stated, but server configuration, permission scope, and connection-failure handling are not described.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Crabtalk This agent | 27 · Major gaps | CLIFree + model costs | ★ 737 | 2mo ago | Rust | — |
| Nausicaa | 72 · Some gaps | CLIFree + model costs | ★ 33 | 1mo ago | TypeScript | ChatGPT · OpenAI API · Claude API |
| TrueForge | 65 · Some gaps | CLIFree + model costs | ★ 6.5k | 2d ago | TypeScript | OpenAI API · Claude API |
| LightAgent | 60 · Some gaps | Library / SDKFree + model costs | ★ 1.2k | 11d ago | Python | OpenAI API |
How does FollowAgents rate this agent?
Why each dimension lost points
The README claims agents declare tools, receive them through separately authorized harnesses, and use RV64 ELF harnesses, but the packet gives no implementation detail or policy evidence, so least privilege receives only partial credit. User confirmation, sensitive-data handling, and rollback are not documented. The README mentions MCP connections and the client snippet shows daemon connectivity, but data flows and controls on external effects are unclear. Cargo.toml lists dependencies without security review or mitigation evidence; the author field gives limited attribution, while publisher identity remains unverified.
The README explicitly warns that the CLI refactor is unfinished and crate layout and APIs are changing. CLAUDE.md says the architecture is currently wrong and docs should not be trusted, weakening self-consistency. The workspace lists crates.io dependencies but does not show a lockfile or availability guarantees. The supplied source offers no assessable evidence about runtime failure messages.
The product is described as a long-running Agent daemon that connects to MCP and can be extended with packages and declared harnesses; these use cases and boundaries remain high-level claims. The packet gives no Agent trigger conditions or scenario limits, and only limited deployment-environment detail, so environment fit is only partly supported.
The README links architecture, installation, configuration, and contribution material, while the Cargo workspace separates apps, crates, harnesses, and libraries, supporting a reasonably clear information structure. Installation notes here are brief and rely on an unavailable linked guide; the CLI refactor also weakens naming and API stability. The README discloses the refactor limitation, but the packet has no FAQ or substantive examples. Apache-2.0 metadata matches the license text and earns full credit. Version 0.0.24 is stated, but no changelog is shown; an author is listed, while maintenance and update responsibility are unclear.
The daemon, MCP connectivity, and extensible harness goals are visible in the README and crate comments, but there are no concrete output examples or user workflows, so usability and marginal value receive only partial support. The refactor status and lack of cost, resource, or deployment information prevent a supported cost-benefit judgment.
Some README architecture claims correspond with the berm and client crate comments, but the key permission claims lack implementation evidence. CLAUDE.md directly disputes the docs and current architecture, limiting corroboration across sources. The README distinguishes product description, architecture claims, and a refactor warning, but the permission claims lack enough evidence to separate assertion from verified behavior with confidence.
- Not found in source: confirmation before actingTurn on (or add) a confirmation step before it acts, and try it in a sandbox or test environment before real data.
- Not found in source: sensitive-data handlingUse dedicated, low-privilege, revocable API keys — never production credentials — and keep secrets out of logs.
- Not found in source: dependency securityPin versions and run a dependency audit (npm audit, pip-audit) before installing; prefer running it in a container.
- Not found in source: rollback or recovery pathBack up first, or work on a git branch or snapshot, so its changes can be undone.
- The README says the CLI refactor is ongoing and internal docs may be out of date; CLAUDE.md also says the current architecture is wrong and docs are untrusted, so treat architecture and permission claims cautiously.
- The packet does not show user confirmation, sensitive-data handling, rollback, or dependency-security processes; do not treat the README's capability-isolation description as implementation-verified.
Maintain this project? Add the FARS badge to your README
[](https://followagents.com/en/agents/crabtalk)<a href="https://followagents.com/en/agents/crabtalk"><img src="https://followagents.com/badge/crabtalk.svg" alt="FARS review"></a>Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
Release notes excerpt
What's Changed * refactor(mcp): extract into standalone crate with hyper backend by @clearloop in https://github.com/crabtalk/crabtalk/pull/181 * refactor(config): single OpenAI-compatible LLM endpoint by @clearloop in https://github.com/crabtalk/crabtalk/pull/182 * feat: introduce crabup as package and service manager by @clearloop in https://github.com/crabtalk/crabtalk/pull/184 * refactor(cron): extract into standalone service by @clearloop in https://github.com/crabtalk/crabtalk/pull/186 …View on GitHub ↗ -
Release notes excerpt
Breaking Changes - Daemon split from TUI (#163) — the daemon is now its own binary (crabtalkd) owning lifecycle/admin/plugin commands. The old crabtalk-node library is renamed to crabtalk; SDK moved to apps/sdk (crabtalk-sdk); CLI moved to apps/tui (crabtalk-tui). SDK consumers must update their imports. …View on GitHub ↗ -
Release notes excerpt
Breaking Changes - Ephemeral delegate agents (#141) — the static worker agent is replaced by on-demand ephemerals. The delegate tool takes name/system_prompt to spawn sub-agents; they're invisible to list_agents. - Session JSONL format (#145) — sessions use a nested HistoryEntry shape. Existing ~/.crabtalk/sessions/*.jsonl files cannot be loaded — delete the directory before upgrading. …View on GitHub ↗
FAQ
Does Crabtalk include a chat interface?
Can it connect to external tool servers?
Which model provider or credentials does it require?
How do applications connect to the daemon?
Are cost, retry, and recovery behavior documented?
Thanks — we use this to improve the page.