# awesome-agent-tools

> Per-capability comparisons of the tools coding agents use. Behaviour is read out of each
> implementation's source; numbers are measured here. Every claim carries where it came from.

Two distinctions matter when quoting anything from this site:

- An *invariant* is something every implementation should hold to, so it is reported as pass/fail.
  A *design choice* is somewhere implementations legitimately differ; it is recorded, never scored.
- An *observation* was read from source. A *measurement* was produced by running a benchmark, and is
  only comparable with another measurement sharing both its metric and its protocol version.

Full database as JSON: https://awesome-agent-tools.com/api/db.json

## Capabilities

### Read — https://awesome-agent-tools.com/tools/read
Nine harnesses read files. What actually differs?
13 tools on file.
- Offset base [design-choice]: 1-indexed (7), 0-indexed (1)
- Announces truncation [invariant]: true (7), unverified (1)
- Binary content never enters context silently [invariant]: true (4), false (1), unverified (4)
- Staleness check before write [invariant]: false (6), true (3)

### Browser — https://awesome-agent-tools.com/tools/browser
I want my agent to drive a browser. Which one?
3 tools on file.

### Web search — https://awesome-agent-tools.com/tools/web-search
Which harnesses can search the web without an add-on?
9 tools on file.

### Code search — https://awesome-agent-tools.com/tools/code-search
How does each harness find code it has not read yet?
24 tools on file.

### Shell — https://awesome-agent-tools.com/tools/shell
Every harness runs shell commands. What differs underneath?
24 tools on file.

### Long-term memory — https://awesome-agent-tools.com/tools/memory
Which agents remember anything between sessions?
5 tools on file.

## Agents

- Claude Code (Anthropic) — https://awesome-agent-tools.com/agents/claude-code — 22 tools
- Codex (OpenAI) — https://awesome-agent-tools.com/agents/codex — 12 tools
- dsh (DeepSeek) — https://awesome-agent-tools.com/agents/dsh — 33 tools
- Grok Build (xAI) — https://awesome-agent-tools.com/agents/grok-build — 21 tools
- Hermes (Nous Research) — https://awesome-agent-tools.com/agents/hermes — 21 tools
- Kimi Code (Moonshot AI) — https://awesome-agent-tools.com/agents/kimi-code — 18 tools
- MiMo Code (Xiaomi) — https://awesome-agent-tools.com/agents/mimo-code — 26 tools
- OpenClaw (openclaw) — https://awesome-agent-tools.com/agents/openclaw — 23 tools
- opencode (anomalyco) — https://awesome-agent-tools.com/agents/opencode — 16 tools
- OpenHands (OpenHands) — https://awesome-agent-tools.com/agents/openhands — 12 tools
- pi (badlogic) — https://awesome-agent-tools.com/agents/pi — 7 tools
- Qwen Code (Alibaba) — https://awesome-agent-tools.com/agents/qwen-code — 12 tools
- ZCode (Z.ai) — https://awesome-agent-tools.com/agents/zcode — 12 tools — no public source, not verified

## Measurements

No measurement runs have been published yet. Benchmark protocols are written down at
https://awesome-agent-tools.com/runs before any numbers exist. Do not quote a measured figure from this site until
one appears there.

## Posts

- Nine read tools, two things you actually need to know — https://awesome-agent-tools.com/posts/nine-read-tools-two-things-to-know — Six of nine coding agents will let you overwrite an edit that landed after the file was read. One counts lines from zero. Neither is visible from the outside until it costs you something.