GLM-5.2 for OpenClaw, Hermes, Pi and Kiro

Evan Brooks

AI Infrastructure Editor

Editorial responsibility: GLM52.ai Editorial Team · Published and editorially reviewed July 26, 2026

How we research and review · Report a correction

Disclosed editorial pen name · AI-generated portrait

A protocol match is not always an allowed product route

OpenClaw and Hermes passed both Z.ai products. Pi passed the metered API, but its Coding Plan config was only parsed offline because Pi is absent from Z.ai’s current supported-tool list. Kiro exposes no arbitrary provider URL or external API-key field, so neither route was called.

OpenClaw 2026.7.1-2 and Hermes Agent 0.19.0 completed GLM-5.2 response and tool probes through both Coding Plan and pay-as-you-go API access. Pi 0.82.1 completed the metered route but remains outside the current Coding Plan support list. Kiro CLI 2.14.2 could select only a Kiro-served model in its custom-agent schema; it could not express either external Z.ai route.

This page covers both requested scenarios for all four tools, including the cases where the correct setup instruction is “do not send the request.” It is aimed at broader agent runtimes, not only coding CLIs. Start them in a disposable, least-privilege environment before allowing shell, browser, messaging, or credential tools.

In this guide

1. Compare all eight route outcomes
2. Configure OpenClaw
3. Configure Hermes Agent
4. Configure Pi coding agent
5. Understand the Kiro limitation
6. Reproduce the response and tool evidence
7. Troubleshoot provider and policy failures
8. Sources and reproducibility boundary

Compare all eight route outcomes

Client	Coding Plan	Pay-as-you-go API	Tool evidence
OpenClaw 2026.7.1-2 (0790d9f)	Pass, OPENCLAW_PLAN_OK, 11,707 ms	Pass, OPENCLAW_PAYGO_OK, 10,179 ms	read returned the sentinel
Hermes Agent 0.19.0 / 2026.7.20 (3ef6bbd…)	Pass, HERMES_PLAN_OK, 14,701 ms	Pass, HERMES_PAYGO_OK, 13,226 ms	terminal tool returned the sentinel
Pi coding agent 0.82.1	Policy-blocked, config parsed offline	Pass, PI_PAYGO_OK, 2,507 ms	read passed on metered API
Kiro CLI 2.14.2	Unsupported, no request	Unsupported, no request	Not applicable

OpenClaw’s plan response reported 13,650 input tokens and its tool run reached 22,318 tokens on the final call because the runtime bootstrapped a workspace and supplied a large default prompt/tool schema. Pi’s metered response reported 380 input and 6 output tokens. Those are client-overhead observations, not a fair model-token benchmark.

Configure OpenClaw

OpenClaw uses a provider catalog plus an agent default model. The two complete files below keep distinct provider IDs, which makes the commercial route visible in logs and avoids a stale merged endpoint.

OpenClaw with Coding Plan

openclaw.json — Coding Plan

{
  "agents": {
    "defaults": {
      "model": { "primary": "zai-plan/glm-5.2" },
      "models": {
        "zai-plan/glm-5.2": {
          "alias": "GLM-5.2 Coding Plan",
          "params": {
            "extra_body": { "thinking": { "type": "disabled" } }
          }
        }
      },
      "timeoutSeconds": 300
    }
  },
  "models": {
    "mode": "merge",
    "providers": {
      "zai-plan": {
        "baseUrl": "https://api.z.ai/api/coding/paas/v4",
        "apiKey": "${ZAI_API_KEY}",
        "api": "openai-completions",
        "timeoutSeconds": 300,
        "models": [{
          "id": "glm-5.2",
          "name": "GLM-5.2 via Coding Plan",
          "reasoning": true,
          "input": ["text"],
          "contextWindow": 1048576,
          "maxTokens": 131072
        }]
      }
    }
  }
}

OpenClaw with pay-as-you-go API

openclaw.json — metered API

{
  "agents": {
    "defaults": {
      "model": { "primary": "zai-api/glm-5.2" },
      "models": {
        "zai-api/glm-5.2": {
          "alias": "GLM-5.2 metered API",
          "params": {
            "extra_body": { "thinking": { "type": "disabled" } }
          }
        }
      },
      "timeoutSeconds": 300
    }
  },
  "models": {
    "mode": "merge",
    "providers": {
      "zai-api": {
        "baseUrl": "https://api.z.ai/api/paas/v4",
        "apiKey": "${ZAI_API_KEY}",
        "api": "openai-completions",
        "timeoutSeconds": 300,
        "models": [{
          "id": "glm-5.2",
          "name": "GLM-5.2 via pay-as-you-go API",
          "reasoning": true,
          "input": ["text"],
          "contextWindow": 1048576,
          "maxTokens": 131072
        }]
      }
    }
  }
}

The test supplied an explicit session key so separate probes did not reuse a previous conversation. Do the same when comparing routes: a shared session can carry cached context, tool state, and instructions across what looks like a clean test.

Configure Hermes Agent

Hermes Agent accepts a custom Chat Completions provider. The pinned test used release 2026.7.20 and source commit 3ef6bbd201263d354fd83ec55b3c306ded2eb72a.

Hermes with Coding Plan

config.yaml — Coding Plan

custom_providers:
  - name: zai-coding-plan
    base_url: https://api.z.ai/api/coding/paas/v4
    key_env: ZAI_API_KEY
    api_mode: chat_completions
    model: glm-5.2
    context_length: 1048576
    extra_body:
      thinking:
        type: disabled

model:
  default: glm-5.2
  provider: custom:zai-coding-plan

agent:
  max_iterations: 2

Hermes with pay-as-you-go API

config.yaml — metered API

custom_providers:
  - name: zai-paygo
    base_url: https://api.z.ai/api/paas/v4
    key_env: ZAI_API_KEY
    api_mode: chat_completions
    model: glm-5.2
    context_length: 1048576
    extra_body:
      thinking:
        type: disabled

model:
  default: glm-5.2
  provider: custom:zai-paygo

agent:
  max_iterations: 2

The two-iteration cap bounded the experiment. Raise autonomy only after a response probe and a harmless terminal read pass. A general agent can call more consequential tools than a code autocomplete client, so provider compatibility is only the first safety gate.

Configure Pi coding agent

Pi 0.82.1 contains a Z.ai-compatible provider shape. The global metered route completed both probes. However, Z.ai’s current Coding Plan list does not name Pi, so we did not use subscription quota from this client.

Pi with Coding Plan: parse only, do not call

This block documents the evaluated technical shape and the policy stop. It is not a working recommendation:

models.json — Coding Plan policy-only record

{
  "providers": {
    "zai-coding-plan-policy-blocked": {
      "baseUrl": "https://api.z.ai/api/coding/paas/v4",
      "api": "openai-completions",
      "apiKey": "$ZAI_API_KEY",
      "compat": {
        "supportsDeveloperRole": false,
        "supportsReasoningEffort": false,
        "thinkingFormat": "zai"
      },
      "models": [{
        "id": "glm-5.2",
        "name": "GLM-5.2 Coding Plan — Pi is not on the supported list",
        "reasoning": true,
        "input": ["text"],
        "contextWindow": 1048576,
        "maxTokens": 131072
      }]
    }
  }
}

The config parsed offline. No plan request was sent, so it must not be labeled “tested working.”

Pi with pay-as-you-go API

models.json — metered API

{
  "providers": {
    "zai-api": {
      "baseUrl": "https://api.z.ai/api/paas/v4",
      "api": "openai-completions",
      "apiKey": "$ZAI_API_KEY",
      "compat": {
        "supportsDeveloperRole": false,
        "supportsReasoningEffort": false,
        "thinkingFormat": "zai"
      },
      "models": [{
        "id": "glm-5.2",
        "name": "GLM-5.2 via Z.ai metered API",
        "reasoning": true,
        "input": ["text"],
        "contextWindow": 1048576,
        "maxTokens": 131072
      }]
    }
  }
}

The metered response returned PI_PAYGO_OK in 2,507 ms with 380 input and 6 output tokens. Its read tool then returned the fixture token; the first and final calls reported 720 and 789 tokens respectively.

Understand the Kiro limitation

Kiro’s custom-agent configuration reference supports a Kiro model selector, prompt, resources, tools, permissions, hooks, and MCP servers. In Kiro CLI 2.14.2, it did not expose an arbitrary model-provider base URL or external API-key field.

The maximum honest diagnostic config was:

glm-5-2-check.json — negative schema probe

{
  "name": "glm-5-2-check",
  "description": "Negative compatibility probe for Kiro CLI 2.14.2",
  "prompt": "Return a short connection status without changing files.",
  "tools": ["read"],
  "model": "glm-5.2"
}

This selects a model name inside Kiro’s own supported model system; it does not connect Z.ai. The validator requested a Kiro login and did not finish within the bounded 12-second probe. There was no field in which to place either:

• Coding Plan: https://api.z.ai/api/coding/paas/v4, or
• pay-as-you-go: https://api.z.ai/api/paas/v4.

Therefore both scenarios are unsupported, not merely untested. Do not paste a Z.ai key into an MCP environment and claim the Kiro agent itself is using GLM-5.2; that would configure one tool, not the agent’s primary model.

Reproduce the response and tool evidence

Use a different token on every run:

Disposable verification fixture

printf 'GLM52_TOOL_%s\n' "$(date +%s)" > agent-tool-probe.txt

For OpenClaw, require its read tool and start with a new session key. For Hermes, name the terminal command and limit the iteration count. For Pi, require read on the metered profile. Check the emitted tool event, path, returned content, final response, and error channel—not just exit status.

Client	Exact response evidence	Tool evidence	What remains unproven
OpenClaw	Both routes returned their route-specific token	read passed	lean prompts, long-session stability, every channel/tool
Hermes	Both routes returned their route-specific token	terminal read passed	higher iteration counts and external extensions
Pi	Metered route returned PI_PAYGO_OK	read passed	Coding Plan eligibility
Kiro	None; no Z.ai call	Not applicable	any direct external provider support

Delete the fixture and temporary agent workspace after the run. Preserve only sanitized version, config, endpoint class, timestamps, response token, tool event, and error evidence.

Troubleshoot provider and policy failures

OpenClaw sends an unexpectedly huge first request

Inspect bootstrapped workspace files, default instructions, memory, enabled channels, and tool schemas. Our first tool probe ended at 22,318 tokens. A large startup request can consume quota or obscure a simple endpoint test even when no repository is attached.

Hermes or Pi returns a model-format error

Keep api_mode or api on Chat Completions and retain the Z.ai compatibility flags from the pinned config. Start with reasoning disabled or bounded, then add optional fields one at a time. Do not mix a Responses-specific body into these Chat routes.

A plan key works in another client but not Pi

Stop. The question is not only authentication. Pi is absent from the current supported-tool list, so do not route around the boundary. Use a listed client or ask Z.ai for explicit support.

Kiro accepts the model name but uses a different model

That is expected from the schema boundary. A name is not a provider connection. Require a configurable endpoint, credential source, protocol, and network evidence before claiming that Kiro itself reached Z.ai.

Plain text works but the tool does not

Reduce the agent to one tool and one exact instruction, verify permissions, and inspect arguments. General agents often carry memory, hooks, extensions, or routing logic that can fail after transport has already succeeded.

Match the subscription to a documented client

OpenClaw and Hermes users should compare their two passing routes against real prompt overhead and usage. Pi users should choose metered API access unless Z.ai explicitly adds Coding Plan support. Kiro users should wait for a provider endpoint rather than forcing a misleading configuration.

Review Z.ai's current agent support →

Affiliate disclosure: This is an affiliate and referral link. GLM52.ai may receive promotional credit after a qualifying purchase at no added charge from us. Z.ai controls supported clients, prices, quotas, and eligibility; the link does not change Pi's policy block or Kiro's unsupported result.

Sources and reproducibility boundary

Sources checked July 26, 2026:

• Z.ai Coding Plan quick start, other-tools page, and API quick start;
• OpenClaw repository;
• Hermes Agent repository;
• Pi coding-agent package;
• Kiro custom-agent configuration reference.

All runnable probes used disposable non-root Docker environments, pinned packages or commits, read-only root filesystems, and no writable host mount. A pass is one dated compatibility observation. Recheck the supported-tool policy and rerun both access scenarios after any provider or client release.

About the editorial role

Evan Brooks

AI Infrastructure Editor

Evan Brooks serves as GLM52.ai’s disclosed editorial pen name for infrastructure, deployment, and training guides. The role examines checkpoint size, memory math, provider limits, implementation evidence, and operational risk. The GLM52.ai Editorial Team produces and fact-checks every article under this byline.

The GLM52.ai Editorial Team accepts responsibility for this article. Read our editorial policy or request a correction.

An AI image tool created this portrait. It does not depict a real individual.