TokenMix Research Lab · 2026-07-22

Grok Build Open Source 2026: Apache 2.0, Local Models, Privacy

Grok Build Open Source 2026: Apache 2.0, Local Models, Privacy

Last Updated: 2026-07-22 Author: TokenMix Research Lab Data verified: 2026-07-22 against the SpaceXAI announcement, Grok Build GitHub repository at commit 3af4d5d, official configuration and enterprise documentation, xAI API pricing, and the original wire-level privacy report.

SpaceXAI open-sourced the Grok Build coding-agent harness under Apache 2.0. It did not release Grok 4.5 model weights. You can compile, audit, fork, and connect the agent to local inference, but the default cloud workflow still requires separate privacy and network controls.

The July 15 release includes the Rust agent loop, file and shell tools, terminal UI, inline diff viewer, skills, plugins, hooks, MCP support, and subagent runtime. The repository says first-party code is Apache 2.0, while vendored components retain their original licenses. Our July 22 source audit checked 2,847 tracked files at commit 3af4d5d39897855bdcc74f23e690024a5dc05573. We found no current source string for the older codebase_upload or git bundle path, but the tree still contains session-trace upload, GCS storage, telemetry, and remote-settings infrastructure. That distinction matters because a reproducible test of Grok Build 0.2.93 found full-repository uploads before SpaceXAI disabled that behavior server-side. The current source is more auditable; it is not proof that every default execution path is offline. SpaceXAI's announcement, the GitHub repository, and the wire-level report establish those separate facts.

Table of Contents

Quick Verdict

Grok Build is genuinely open-source software, but "Grok is open source" is still a misleading summary.

Claim Status Source
Grok Build source code is public Confirmed SpaceXAI announcement
First-party code uses Apache 2.0 Confirmed Repository license
Grok 4.5 model weights were released False The release contains the harness, not the model weights
Developers can connect a custom or local model Confirmed Grok Build documentation
The default Grok 4.5 inference path is free False Grok 4.5 API tokens remain billable
Open source automatically means no network traffic False Authentication, inference, telemetry, trace upload, and session sync are separate paths
Version 0.2.93 uploaded full repositories in an independent test Confirmed for the tested setup Wire-level report
That test proved SpaceXAI trained on the code False The researcher explicitly limited the finding to transmission and storage
The old whole-repository upload remains active today Unresolved A current-binary wire test is still required
Open-sourcing was caused by the privacy disclosure Likely The timing is close, but SpaceXAI did not state that reason

The practical verdict is direct: use the repository as an auditable agent framework. Do not treat it as an open-weight Grok release or as a privacy guarantee.

What SpaceXAI Actually Open-Sourced

The release covers the software around the model: the agent runtime, terminal interface, tools, and extension system.

Repository area What it contains Why it matters Status
xai-grok-shell Agent runtime, headless mode, session handling Controls the model-tool loop Confirmed
xai-grok-tools Terminal, file editing, and search tools Defines what the agent can do on a workstation Confirmed
xai-grok-pager Full-screen terminal UI Makes the agent usable interactively Confirmed
xai-grok-workspace Filesystem, VCS, execution, checkpoints Sits closest to repository data Confirmed
Skills and plugins Extension discovery and loading Enables reusable workflows Confirmed
MCP support Local and remote tool servers Extends data and action surfaces Confirmed
Subagents Delegated agent tasks Supports parallel or specialized work Confirmed
Grok 4.5 weights Not included The intelligence layer remains separate Confirmed absence

The repository describes itself as a periodic sync from SpaceXAI's internal monorepo. At the audited commit, the public tree contained 2,847 tracked files and identified Rust as the primary implementation language. The README also says external contributions are not currently accepted. Users can inspect and fork the code, but "open source" does not yet mean a conventional upstream community workflow where outside pull requests are merged.

SpaceXAI publishes prebuilt binaries for macOS, Linux, and Windows. The source-build instructions list macOS and Linux as supported build hosts and describe Windows source builds as best-effort and not currently tested from the public tree. That is a packaging limitation, not a license limitation.

Apache 2.0 License and Commercial Use

Apache 2.0 permits commercial use, modification, and redistribution of Grok Build's first-party code, subject to its notice and license obligations.

Use Apache 2.0 position Practical requirement Status
Internal company deployment Allowed Keep required notices Confirmed
Fork and modify the agent Allowed Document material changes where required Confirmed
Redistribute binaries Allowed Include license and notices Confirmed
Build a commercial product Allowed Review trademarks and third-party licenses separately Confirmed
Claim Grok 4.5 is included Not supported Obtain model access separately Confirmed
Re-license third-party code as Apache 2.0 Not automatically allowed Preserve each component's original license Confirmed
Submit changes upstream Currently not accepted Maintain a fork or wait for policy change Confirmed

The repository README explicitly separates first-party code from third-party and vendored code. Its notices include components derived from OpenAI Codex and OpenCode implementations. A commercial fork therefore needs a license inventory, not just a copy of the root Apache file.

Apache 2.0 also does not transfer SpaceXAI trademarks or guarantee continued access to Grok APIs. The harness license and model service agreement remain two different contracts.

Grok 4.5 Is Not Open Source

Grok 4.5 remains an API model; the open repository does not contain its weights, training code, or training data.

Layer Open in this release? Can run without SpaceXAI? Status
Grok Build agent loop Yes Yes, after compiling or using the binary Confirmed
Terminal UI Yes Yes Confirmed
File and shell tools Yes Yes Confirmed
Skills, plugins, MCP, hooks Yes Yes Confirmed
Grok 4.5 weights No No local Grok 4.5 inference from this release Confirmed
Grok 4.5 API Service access Requires an API key or eligible product access Confirmed
A different local model Bring your own Yes, if a compatible inference endpoint is available Confirmed

Official xAI documentation lists grok-4.5 with a 500K context window and current short-context prices of $2 per million input tokens, $0.30 per million cached tokens, and $6 per million output tokens. Prompts at or above 200K tokens use the published long-context rates of $4 input, $0.60 cached input, and $12 output. See the current xAI pricing page and our Grok 4.5 review.

Calling this a "Grok model open-source release" creates three bad assumptions:

  1. It suggests the Grok 4.5 weights can be downloaded.
  2. It suggests model inference has no token bill.
  3. It suggests the entire stack becomes offline automatically.

All three are false.

Local Models and Local-First Architecture

Grok Build can use a local model because its model layer accepts a configurable model ID, base URL, display name, and credential environment variable.

The official configuration schema looks like this:

[model.my-model]
model = "model-id"
base_url = "https://api.example.com/v1"
name = "Display Name"
env_key = "API_KEY"

[models]
default = "my-model"

For a local OpenAI-compatible inference service, the same shape can point to localhost:

[models]
default = "local-coder"

[model.local-coder]
model = "your-local-model-id"
base_url = "http://127.0.0.1:8000/v1"
name = "Local Coder"
env_key = "LOCAL_MODEL_KEY"

The exact model ID, authentication behavior, tool-call format, and context limit depend on the local inference server. The example is a configuration pattern, not a claim that every open model can execute Grok Build's agent loop correctly.

Architecture Agent location Model location Main benefit Main caveat
Standard Grok Build Local CLI SpaceXAI cloud Fastest setup and Grok 4.5 access Code context needed by the model leaves the machine
Local harness + xAI API key Local CLI SpaceXAI API Direct billing and model control Still cloud inference
Local harness + company gateway Local CLI Company-controlled route Central policy, logs, fallback Gateway still forwards prompts upstream
Local harness + local inference Local CLI Local workstation or private server Maximum model-path control Hardware, compatibility, and quality are your responsibility
Local harness + private cloud model Local CLI Private VPC endpoint Enterprise isolation and scale Infrastructure and operational cost

Run grok inspect after editing the configuration. It reports the effective model, instructions, skills, plugins, hooks, and MCP servers detected for the current workspace.

grok inspect
grok -p "Explain the repository architecture" -m local-coder

Local inference does not make every tool local. Remote MCP servers, web search, authentication, updates, session sharing, and plugins may create their own connections. A privacy review must follow the whole graph, not only the model URL.

For multi-provider routing, use an explicit gateway policy rather than changing endpoints manually. The AI API gateway guide covers routing, fallback, and observability patterns.

The Privacy Catch

Open-sourcing makes Grok Build inspectable, but it arrived after a verified privacy incident involving an older binary.

An independent researcher tested Grok Build 0.2.93 with controlled repositories and intercepted the tool's own traffic. The report found two separate channels: model requests containing files the agent read, and storage uploads containing a repository snapshot. In one preserved test, approximately 5.10 GiB moved through storage requests while model-turn traffic totaled about 192 KB. A captured Git bundle could reconstruct a file the agent had been told not to read, together with Git history.

The researcher explicitly did not prove that SpaceXAI trained on the uploaded data. The evidence established transmission, server acceptance, and storage behavior for the tested version and account. Axios independently reported that uploads appeared to stop after disclosure and that SpaceXAI said previously uploaded data would be deleted.

Date Event What is established Status
2026-07-08 Research test used Grok Build 0.2.93 Tested binary and version are identified Confirmed
2026-07-12 Automatic whole-codebase upload was reportedly disabled Independent reporting and researcher update agree Confirmed by reporting
2026-07-14 Researcher updated the disclosure Old upload disabled server-side; deletion not independently verified Confirmed
2026-07-15 Grok Build source published Agent harness released under Apache 2.0 Confirmed
2026-07-21 Public repository synced again Our audited commit date Confirmed
2026-07-22 Current-binary wire behavior No fresh reproduction in this review Unresolved

The chronology supports a Likely inference that transparency became more urgent after the disclosure. It does not prove SpaceXAI open-sourced the project because of the incident. The official announcement presents reliability, extensibility, and local-first deployment as its reasons.

What the Current Source Code Shows

The July 21 source tree removes evidence of the named whole-codebase path, but retains configurable upload and telemetry systems.

We searched the audited commit for codebase_upload, upload_codebase, git_bundle, and git bundle; none appeared in tracked Rust, Markdown, or TOML files. That is useful evidence that the old path is not present under those names. It is not a network test and cannot prove equivalent behavior is impossible through a differently named path.

The current tree does contain:

Current-source finding Evidence What it proves Status
No codebase_upload string Repository-wide search at commit 3af4d5d Old named path is absent Confirmed
No git bundle string Repository-wide search Old named implementation is not visible under that phrase Confirmed
Session trace upload exists trace_upload, storage client, GCS utilities Upload capability remains Confirmed
Telemetry fallback is disabled resolve_telemetry_mode() default No-override source fallback is off Confirmed
Remote settings can set telemetry Config resolution order Account/server configuration can affect resolved behavior Confirmed
Trace upload follows telemetry when unset resolve_trace_upload() One control can indirectly enable the other Confirmed
/privacy is separate Current user guide Privacy opt-out does not automatically disable telemetry Confirmed
Current binary never uploads repository data No fresh packet capture Cannot be concluded from static search alone Unresolved

This is the most important technical conclusion in the article:

Open source improves auditability. It does not replace runtime verification.

Privacy-Hardened Configuration

The safest configuration explicitly disables telemetry and trace uploads instead of assuming /privacy covers both.

The current user guide says these controls are independent:

A conservative starting configuration is:

[features]
telemetry = false

[telemetry]
trace_upload = false
mixpanel_enabled = false

[models]
default = "local-coder"

[model.local-coder]
model = "your-local-model-id"
base_url = "http://127.0.0.1:8000/v1"
name = "Local Coder"
env_key = "LOCAL_MODEL_KEY"

The corresponding environment controls are:

export GROK_TELEMETRY_ENABLED=false
export GROK_TELEMETRY_TRACE_UPLOAD=false
grok inspect
grok -p "List the files you would need before reading them" -m local-coder

Environment variables and TOML syntax above come from the current repository documentation. We did not packet-capture this configuration, so "zero outbound traffic" remains a condition to verify, not a promised result.

Control What it changes What it does not guarantee Status
/privacy opt-out Coding-data sharing preference Telemetry and trace upload are separate Confirmed
telemetry = false Product analytics resolution Model prompts still go to the configured model endpoint Confirmed
trace_upload = false Session trace upload setting Plugins and MCP servers may still connect externally Confirmed
Local base_url Model inference route Updates, auth, web tools, and remote sessions may remain Confirmed
Blocking code.grok.com Disables remote session sync/sharing Does not block inference proxy by itself Confirmed
Packet capture Observes actual connections Does not prove future versions behave identically Confirmed

For sensitive repositories:

  1. Start with fake credentials and a canary repository.
  2. Disable telemetry and trace upload explicitly.
  3. Use a local model endpoint.
  4. Block nonessential domains at the network layer.
  5. Capture DNS and HTTPS destinations.
  6. Confirm ignored files, Git history, plugins, and MCP data remain inside the intended boundary.
  7. Repeat after every client update.

The local AI agent SSH workflow covers another important boundary: keeping credentials on the server instead of exposing private keys to a desktop agent.

Cost Math: Open Harness, Paid Inference

The Grok Build license costs $0, but Grok 4.5 inference remains usage-priced.

The calculations below use xAI's July 22 published short-context prices: $2 per million input tokens and $6 per million output tokens. They exclude tool calls, taxes, retries, priority processing, and long-context multipliers.

Monthly workload Input cost Output cost Total Status
1M input + 0.2M output $2.00 $1.20 $3.20 Confirmed rates, calculated
10M input + 2M output $20.00 $12.00 $32.00 Confirmed rates, calculated
100M input + 20M output $200.00 $120.00 $320.00 Confirmed rates, calculated

At 50 coding sessions per workday and 22 workdays, a $0.20 average model bill per session becomes:

50 sessions x 22 days x $0.20 = $220 per developer per month

That $0.20 average is a scenario assumption, not a measured Grok Build average. Teams should replace it with usage exported from their own sessions.

Long prompts change the math. xAI says prompts at or above 200K tokens use $4 per million input and $12 per million output for Grok 4.5.

One request Short-context math Long-context billed math Difference Status
300K input + 50K output $0.90 $1.80 +$0.90 Confirmed rates, calculated
500K input + 100K output $1.60 $3.20 +$1.60 Confirmed rates, calculated

Running a local model removes the xAI token charge but does not make inference free. Hardware purchase, GPU rental, electricity, engineering time, and lower task-completion rates can dominate the result. Without a measured local deployment, assigning a dollar saving would be Speculation.

For current account and model access steps, use the Grok API key guide and verify the live console before budgeting.

Use Case Matrix

Grok Build's open-source release is most valuable to teams that need an auditable agent shell, not to users expecting free Grok 4.5 weights.

Use case Recommendation Why Confidence
Audit agent file and shell behavior Inspect or fork Grok Build Relevant source is now public Confirmed
Use Grok 4.5 with minimal setup Standard Grok Build Official supported path Confirmed
Run an open model locally Grok Build plus local endpoint Custom model configuration exists Confirmed, compatibility requires testing
Regulated source-code environment Fork, pin version, disable uploads, capture traffic Defaults alone are insufficient evidence Likely best practice
Build a commercial coding tool Apache fork plus license review First-party code permits commercial use Confirmed
Accept community pull requests upstream Do not depend on it today Official repository says external contributions are not accepted Confirmed
Assume privacy because source is open Do not proceed Runtime config and network behavior still matter Confirmed
Replace a working coding agent immediately Run a canary first Harness and model quality are separate Likely

Risks and Caveats

The largest risk is collapsing four separate questions - source availability, model weights, runtime traffic, and data retention - into one word: open.

Risk Likelihood Impact Mitigation Status
Calling Grok 4.5 open source High Misleads readers and buyers Say "Grok Build harness" every time Confirmed
Assuming local-first means offline High Unexpected data transfer Use local endpoint and network controls Confirmed
Relying only on /privacy Medium Telemetry or traces may remain configured Disable each control explicitly Confirmed
Applying 0.2.93 findings to every current build Medium Overstates current risk Label tested version and re-test current binary Confirmed caveat
Assuming old behavior is fully fixed Medium False security Packet-capture the deployed version Unresolved
Ignoring third-party licenses Medium Redistribution exposure Audit notices before shipping a fork Confirmed
Expecting upstream contributions High Fork maintenance burden Budget for internal ownership Confirmed
Using a weak local model for agent loops Medium More retries and failed tasks Benchmark task completion, not token price Likely
Letting remote settings override assumptions Medium Configuration drift Pin requirements and inspect effective config Confirmed capability

Final Recommendation

Use Grok Build as an auditable Apache 2.0 agent framework, not as an open-weight Grok release. For ordinary coding, the official Grok 4.5 path is the simplest. For sensitive code, pin the source, choose a controlled model endpoint, disable telemetry and trace upload explicitly, block optional services, and verify the deployed binary on the wire.

FAQ

Is Grok Build open source?

Yes. SpaceXAI published the Grok Build coding-agent harness and terminal UI under Apache 2.0 on July 15, 2026. First-party code can be inspected, modified, redistributed, and used commercially under the license terms.

Is Grok 4.5 open source?

No. The Grok Build repository does not include Grok 4.5 weights, training code, or training data. Grok 4.5 remains available through SpaceXAI's products and API.

Can Grok Build run a local model?

Yes, if the model is exposed through a compatible inference endpoint. Grok Build supports custom model IDs, base URLs, names, and credential environment variables. Tool calling and context behavior still need workload testing.

Does local-first mean Grok Build is fully offline?

No. A local model removes the cloud model path, but authentication, updates, plugins, MCP servers, web tools, telemetry, trace upload, and remote sessions are separate network surfaces. Verify actual traffic before using sensitive repositories.

Can I use Grok Build commercially?

Yes, the first-party code is licensed under Apache 2.0. You must preserve required notices and separately review third-party licenses, trademarks, and any model-service terms.

Did Grok Build upload complete repositories?

An independent wire-level test confirmed that Grok Build 0.2.93 uploaded complete repository data in the tested setup. SpaceXAI reportedly disabled that behavior afterward. This article does not claim the current binary still behaves the same way.

Does /privacy opt-out disable all uploads?

No. Current Grok Build documentation says coding-data privacy, product telemetry, session trace upload, and external OpenTelemetry are independent controls. Disable and verify each relevant path separately.

How much does Grok 4.5 cost with Grok Build?

The harness has no license fee, while API inference is usage-priced. xAI currently lists short-context Grok 4.5 at $2 per million input tokens, $0.30 cached input, and $6 output; prompts at or above 200K use higher long-context rates.

About TokenMix

TokenMix.ai is an AI API relay for developers who need one OpenAI-compatible endpoint across OpenAI, Anthropic, Google, DeepSeek, Qwen, GLM, Grok, and other model families. Check the live model catalog, compare pricing, and read the API docs before routing production workloads.

Sources

Related Articles