Run Orbi with z.ai GLM 5.3 Flash

This is the first article in the provider series: a copyable setup for running your first Orbi Issue with z.ai’s OpenAI-compatible endpoint.

Referral: 👉 Join z.ai now

What you can get for free

As of 2026-09-04, we verified one real Orbi delivery with glm-5.3-flash. That is evidence that the route works, not a promise of a universal free quota. The following are not measured in this article: the account’s exact free allowance, whether a card is required, the exhaustion/reset behavior, rate-limit response, elapsed wall-clock time, and tokens or quota consumed by one task. These vary by account and current z.ai policy. Do not infer them from the model’s 1M-class catalog claims or from this example.

Prerequisites

You need a GitHub repository with Orbi installed and a z.ai API key. Create the key in the z.ai console. Keep it out of Git, Issues, PRs, and journals. The endpoint and API style used here are:

endpoint: https://api.z.ai/api/paas/v4
model: glm-5.3-flash
API: openai-completions
authentication: Authorization: Bearer <key> (Pi sends this from apiKey)

Configure it

Copy the provider template, or create .orbi/pi-providers.json with this complete content:

{
  "providers": {
    "z-ai": {
      "baseUrl": "https://api.z.ai/api/paas/v4",
      "api": "openai-completions",
      "apiKey": "$ZAI_API_KEY",
      "models": [
        {
          "id": "glm-5.3-flash",
          "name": "GLM 5.3 Flash",
          "contextWindow": 131072,
          "maxTokens": 32768
        }
      ]
    }
  }
}

Create the local environment file. Replace the placeholder locally; never commit the resulting file:

mkdir -p .orbi
printf '%s\n' 'ZAI_API_KEY=replace-with-your-key' > .orbi/env
chmod 600 .orbi/env

Select the exact provider and model in orbi.toml:

pi_providers = ".orbi/pi-providers.json"
pi_provider = "z-ai"
pi_model = "glm-5.3-flash"

For a manual tick, systemd is not loading .orbi/env. Export it first, then run Orbi using the normal command from the getting-started guide:

set -a
. .orbi/env
set +a
pi --print "reply with the single word: ok"
orbi --config orbi.toml

Run one real pi --print check before dispatching an Issue. The Runner also fails fast if the JSON, selected model, endpoint, or ZAI_API_KEY is invalid.

The real run

This article reuses a delivery that was already run on 2026-09-04; no new Issue was dispatched for this article:

Issue #303
run ID: 57517638
result: commit b81f77a and PR #304, now merged
implementation: 45 lines, plus 159 lines of tests and synchronized Chinese and English documentation (209 lines total)
session evidence: 148 records, 435 KB, 11 assistant responses, 15 tool calls
elapsed time: not recorded in the supplied evidence
quota/tokens consumed: not measured

A useful detail: #303 fixed a bug where an apiKey written as $VAR in pi_providers did not expand. The provider used to run this example helped fix the configuration path that lets providers like z.ai work. To inspect a live run, follow the journal and the run ID:

journalctl --user -u [email protected] -u [email protected] | grep 57517638

The PR is the delivery boundary. Orbi does not silently retry or rotate to another provider in the same session.

Pitfalls

Confirm the key-expansion fix

PR #304 is merged, so the current main includes the #303 fix. If you are using an older Orbi checkout, $ZAI_API_KEY may not work; update first rather than putting a real key into committed JSON.

`session_pending` does not necessarily mean stuck

Do not judge a run only by how long its phase has stayed unchanged. During session_pending, the model may be working intensively. Check whether the .pi-session/*.jsonl file is still growing, and use the journal’s activity lines. A genuinely stale startup is handled by Orbi’s recovery path; do not invent a retry loop.

Reasoning metadata is deliberately omitted

The working provider template does not guess a provider-specific reasoning or thinking-level mapping. Do not copy Gemini’s thinkingLevelMap into this file without verifying the current Pi and z.ai contracts.

When the allowance runs out

The exact exhaustion message and recovery time are not tested here. Keep the failed run evidence, check the current z.ai response, and switch manually for the next run. There is no automatic fallback. Replace the provider file and key with another tested or explicitly not-tested template, then change both selectors to matching values. For example, the OpenRouter template uses:

pi_providers = ".orbi/pi-providers.json"
pi_provider = "openrouter"
pi_model = "google/gemma-4-31b-it:free"

Set OPENROUTER_API_KEY, validate the exact entry, and run one real pi --print check before dispatching. That alternative’s availability and quota are not tested in this article. Suggested post (use the real evidence, not a quota promise):

We ran a real Orbi Issue with z.ai GLM 5.3 Flash: 45 lines of implementation, 159 lines of tests, and a PR opened hands-free. The run produced 148 session records. Configure it here: https://orbi.build/apply

For the complete provider catalog and current status, see the provider guide. For the Chinese version of this article, see 用 z.ai GLM 5.3 Flash 跑 Orbi.

Getting Started

Core Concepts

Guides

Optional Components

Releases

Zai glm 5 3 flash

Run Orbi with z.ai GLM 5.3 Flash

What you can get for free

Prerequisites

Configure it

The real run

Pitfalls

Confirm the key-expansion fix

`session_pending` does not necessarily mean stuck

Reasoning metadata is deliberately omitted

When the allowance runs out

​Run Orbi with z.ai GLM 5.3 Flash

​What you can get for free

​Prerequisites

​Configure it

​The real run

​Pitfalls

​Confirm the key-expansion fix

​session_pending does not necessarily mean stuck

​Reasoning metadata is deliberately omitted

​When the allowance runs out

​Share after publishing

Run Orbi with z.ai GLM 5.3 Flash

What you can get for free

Prerequisites

Configure it

The real run

Pitfalls

Confirm the key-expansion fix

`session_pending` does not necessarily mean stuck

Reasoning metadata is deliberately omitted

When the allowance runs out

Share after publishing