Native Kiro selectors and slash commands inside Zed's Agent Panel.
Kiro
Agent Default ▾
Model Claude Opus 5 ▾
Effort Max ▾
Zed and Kiro both speak the Agent Client Protocol, but different generations of it.
Kiro CLI 2.21.0 exposes model selection through session/set_model and a models
block on session/new — an unstable API that was removed from ACP and removed
from Zed in zed#58308. Current Zed
drives models, modes and reasoning effort exclusively through
session/set_config_option, which Kiro answers with -32601 Method not found.
Neither side is broken. They just cannot see each other. The visible symptoms are zed#59169 (no model picker), kiro#9998 (reasoning effort unreachable) and kiro#10034.
This bridge sits between them and translates. Zed sees clean, standard ACP; Kiro keeps using the protocol it already implements.
Kiro CLI 2.28's opt-in CLI V3 engine speaks config options natively; the bridge has a separate V3 adapter that adds validation, command routes and a safe default, while Zed sees the same selectors and commands on either engine.
- Native model picker with Kiro's real models, discovered dynamically, showing
each model's credit multiplier. Switch models mid-thread — no restart, no new
conversation. Newly released models are detected in the background and appear in
new threads automatically;
/restart-kiroloads them into an open thread. - Native reasoning-effort picker whose options are re-queried per model. GPT models
offer
none; Claude models do not;autohas no effort axis, so the selector is withdrawn and you are told why. - Native agent picker including Plan, Guide and your own custom agents from
~/.kiro/agents/. - Slash commands — type
/and Kiro's commands appear, including all your skills from~/.kiro/skills/and.kiro/skills/, which Kiro does not publish over ACP at all. - Real context usage in Zed's context ring, using measured token counts rather than numbers inferred from a percentage.
- Tool calls, permission prompts, cancellation and images, with two Kiro path defects corrected so diffs and file links point at the right files.
- Sign in from inside Zed. Kiro advertises no ACP auth method at all, so a
signed-out user has no way in; the bridge adds ACP Terminal Auth backed by
kiro-cli login.
- Node.js ≥ 22
- Kiro CLI, already signed in (
kiro-cli --versionshould work) - Zed with external agent support
The bridge holds no credentials. It uses your existing authenticated Kiro CLI.
npm install -g kiro-acp-bridgeOr without installing anything globally, let Zed fetch it on demand — see docs/zed-setup.md.
Then in Zed's settings.json:
{
"agent_servers": {
"Kiro": {
"type": "custom",
"command": "kiro-acp-bridge",
"args": [],
"env": {}
}
}
}Open the Agent Panel, pick Kiro as the agent, and the three selectors appear.
Full options, including npx usage and remote projects, are in
docs/zed-setup.md.
Zed ──standard ACP v1──▶ kiro-acp-bridge ──legacy ACP + _kiro.dev/*──▶ kiro-cli acp
One process, two ACP connections: an agent northbound to Zed, a client southbound to Kiro. All Kiro-specific vocabulary stays on the southbound side.
The three translations that matter:
| Zed asks for | Bridge sends to Kiro |
|---|---|
set_config_option model |
session/set_model |
set_config_option effort |
commands/execute {effort, {level}} |
set_config_option agent |
session/set_mode |
See docs/architecture.md for module layout and docs/architecture-decision.md for why a bridge rather than a fork.
| Component | Verified |
|---|---|
| Kiro CLI | 2.28.0, engine v2 (default) and CLI V3 (opt-in) |
| ACP | protocol version 1 (stable) |
| SDK | @agentclientprotocol/sdk 1.7.0 |
The CLI V3 engine is supported with KIRO_BRIDGE_AGENT_ENGINE=v3 (or auto, which falls
back to v2). v2 remains the default because V3 does not yet report token counts and lacks
some v2 commands; see compatibility.
Feature detection is used throughout rather than version checks, so most Kiro updates need no change here. See docs/compatibility.md.
- No slash-command argument autocomplete. ACP v1 has no completion mechanism, and
Zed reads
AvailableCommand.inputas a boolean while discarding the hint text. Not fixable in a bridge. - Image support is per-model, but ACP advertises it per-agent. Sending an image to a non-vision model surfaces Kiro's error rather than silently degrading.
- MCP OAuth is implemented but not verified end-to-end — no OAuth-requiring MCP server was available to test against.
/spawnand/rewindwork as commands but have no richer Zed UI, because ACP has no surface for parallel sessions or checkpoint selection.
Full list with reasons: docs/capability-matrix.md.
"Kiro CLI was not found" — the message lists every path searched and how to fix it.
Set KIRO_CLI_PATH in the env block of your Zed agent config.
Selectors missing — confirm you selected the bridge, not kiro-cli directly.
Check with dev: open acp logs in Zed's command palette: session/new should return
a configOptions array.
Anything else — enable tracing and read the frames:
KIRO_BRIDGE_TRACE=1 KIRO_BRIDGE_TRACE_FILE=/tmp/kiro-bridge.log kiro-acp-bridgeMore in docs/debugging.md.
- No telemetry. No network calls of its own — it only speaks to your local
kiro-cli. - No credentials stored or read. Kiro owns authentication.
- Prompt content is not logged, even with tracing on.
KIRO_BRIDGE_TRACE_CONTENT=1opts in explicitly; the resulting file will contain your prompts. - Secret-bearing keys (tokens, API keys, authorization headers, client secrets) are redacted from all diagnostics.
- OAuth URLs are never logged and never opened by the bridge — they go to Zed, which asks you first.
- Kiro's permission model is untouched: the bridge never auto-approves a tool call and never adds an option Kiro did not offer.
- Subprocesses are spawned with a fixed argument array and
shell: false, so there is no shell interpolation path.
npm install
npm run build
npm test # unit + fake-Kiro integration tests (no kiro-cli needed)
node scripts/handshake.mjs # real Kiro handshake, proves clean teardownEnd-to-end scripts under scripts/ drive the built bridge through a simulated Zed
against real Kiro. They run Kiro under a sandbox HOME (scripts/lib/sandbox.mjs) so
they never touch ~/.kiro, and keep credit spend negligible.
node scripts/e2e-engines.mjs # v2 + CLI V3, /restart-kiro, ~/.kiro leak check
node scripts/e2e-spine.mjs # streaming, tools, permissions, paths, cancel
node scripts/e2e-config.mjs # model / effort / agent selectors
node scripts/e2e-commands.mjs # slash commands, skills, state sync
node scripts/e2e-usage.mjs # context usage, credits, session list
node scripts/e2e-failures.mjs # failure paths, redaction, crash, imagesApache-2.0. See docs/research.md for licence findings on the projects studied as references — no code was copied from any of them.