Veecle

Chiplab is itself the tool an agent calls. Its MCP server exposes ask (a knowledge base assistant) and run (firmware on a virtual board), and the customer's agent host decides what else the agent reaches. No connector catalog, custom tool route or action framework of Chiplab's own lets agents take authenticated action in outside systems.

Renode, the simulator behind run, is Chiplab's execution backend rather than an integration the agent acts through. The Enterprise tier's custom chip and tool integrations add targets to Chiplab rather than systems the agent reaches. Partner environment models appear only in the planned simulate tool.

Sourcedocs.veecle.ai/tools/overviewread 2026-09-23

No multi step execution of Chiplab's own ships today. The run reference says a run may be a single call or a short sequence (issue an upload slot, upload the artifact, trigger the run) that the customer's agent discovers and sequences itself. The orchestrator tool, automate, which would chain build, run, test, bench, simulate and ask, is on the roadmap for 2027 and later.

Sourceveecle.ai/roadmapread 2026-09-23

The ask tool routes a question to an assistant scoped to Chiplab usage topics. It is grounded in Chiplab's documentation and a shared knowledge corpus of chip level behavior observed across the platform's runs, indexed by chip family, failure pattern and board configuration.

No retrieval structure over the customer's own corpus is documented. The corpus is pooled across every user and, by Chiplab's own terms, never collects the customer's firmware or binaries, so it is Chiplab's own knowledge product, much like an index of the public web.

Sourcedocs.veecle.ai/tools/askread 2026-09-23

No review, approval or pause for a person step is documented. The agent uploads, runs and reads results on its own, and the reference pages tell the buyer they need not call anything themselves. The dashboard's revoke control on a connected agent session is access control, not a review gate.

Sourcedocs.veecle.ai/tools/runread 2026-09-23

Neither identity integration nor a compliance attestation is documented. Chiplab offers no sign in through the customer's own identity provider and no named access model with roles or permissions. Agents authenticate to Veecle's own accounts through a browser based OAuth 2.0 flow. The API Keys page lists authorized agent sessions with first connected and last active times and a revoke control, which is credential management on a single account rather than a view of who in the buyer's organization can do what.

On compliance, the site footer carries no certification badge or trust link, and no trust center is published. The privacy notice is a GDPR controller statement naming processors (Supabase in Frankfurt or Ireland, Cloudflare, Stripe) and a retention schedule, which is disclosure rather than attestation.

Sourcedocs.veecle.ai/platform/api-keysread 2026-09-23

The agent's requests are recorded. The Usage dashboard loads the account's full request history for the current billing period, with aggregate totals and a recent window for confirming a run went through, and the API Keys page shows each agent session's first connected and last active times. That is a record of what the agent did on Chiplab, kept for a billing period.

Inspecting prompts, tool calls and outputs step by step is left to the agent's own host, and no audit log separate from usage, export to a monitoring system or retention by plan is documented. The console output and fault diagnostics a run returns describe the firmware's behavior, not the agent's.

Sourcedocs.veecle.ai/platform/usageread 2026-09-23

Runs return synchronously, bounded to a fixed amount of virtual CPU time, and no state carried from one run into the next is documented. The shared knowledge corpus each run feeds is absorbed, platform wide learning rather than a memory the buyer's agent reads, reviews or edits. A Chiplab blog post describes each run as an addressable object with an identity of its own. Nothing on the product pages documents resuming a held chip from that handle.

Sourcedocs.veecle.ai/introductionread 2026-09-23

Chiplab is a hosted MCP service, and the pricing page's Enterprise tier names on premises deployment alongside volume commits, custom chip and tool integrations and SLAs. So customers can run it in their own environment.

On premises deployment is one line on a custom tier with no deployment documentation behind it, so how an on premises install is delivered is not described. The privacy notice places the hosted service's database and authentication with Supabase in the EU (Frankfurt or Ireland), which is a fact about Chiplab's own infrastructure rather than a region the buyer selects.

Sourceveecle.ai/pricingread 2026-09-23

Chiplab ships ready made starting points. Its public examples repository carries a ready to run firmware example for every supported board across five frameworks (bare metal and Embassy Rust, Zephyr, FreeRTOS and ThreadX), each with an AGENTS.md telling a coding agent how to build, upload and run it. These are examples a developer's agent builds from, not a catalog of agents a buyer adopts.

Sourcegithub.com/veecle/chiplabread 2026-09-23

Work starts only when the customer's agent calls. The agent overview says the coding agent discovers the tools and calls them when a message involves firmware simulation. No event, schedule, webhook or inbound queue of Chiplab's own starts work. The docs' suggested rule, adding 'test them on Chiplab before reporting back' to CLAUDE.md, AGENTS.md or a Cursor rule, lives in the customer's agent. Per commit testing belongs to the test tool, which the roadmap lists as in progress.

Sourcedocs.veecle.ai/agents/overviewread 2026-09-23

Chiplab runs no model the buyer chooses. Any MCP capable coding agent calls it (Cursor, Claude Code, Codex, VS Code, OpenCode and others), and the assistant behind ask names no model and offers no selection. As an MCP server, Chiplab sits on the other side of the model choice: the customer picks the assistant that calls it.

Sourcedocs.veecle.ai/agents/overviewread 2026-09-23

The only way in from outside is Chiplab's own HTTP MCP server. The endpoint is published (chiplab.veecle.ai/mcp), the auth scheme is published (a browser based OAuth 2.0 flow with automatic client registration, with setup guides for six clients), and the tools are enumerated (ask and run, with parameters and response fields).

What is missing is a stable contract. The tools overview describes its tool names, parameters and response shapes as illustrative rather than a frozen API contract, and the agent overview presents them as calls the agent makes, with no manual API calls. Two tools Veecle itself calls illustrative are a sample, not the product's interface.

No REST API or SDK is documented, and the API Keys page manages OAuth agent sessions rather than API keys. The home page's 'single API' is the MCP endpoint.

Sourcedocs.veecle.ai/tools/overviewread 2026-09-23

The customer's own artifact is what gets tested. The run tool executes the firmware the agent built on a chip accurate virtual board (same binary, peripherals and interrupt timing as the physical part, per the run reference) and returns console output with simulator diagnostics for faults. Every agent setup guide suggests gating changes on it ('after making firmware changes, test them on Chiplab before reporting back'). Because it tests the customer's code, this is more than a sandbox that supplies compute.

The test tool (unit and integration tests, coverage, regression tracking) is on the roadmap as in progress, and the introduction's 'pass/fail status' is not among the run reference's response fields (run ID, stdout, stderr). What ships today is console output and fault diagnostics read after a run.

Sourcedocs.veecle.ai/tools/runread 2026-09-23

Agents reach Chiplab through MCP tool calls that upload an ELF and run it on a virtual board. No browser or desktop the agent operates, and no screenshot, click or typing surface, is documented. The virtual chip is an execution target driven through tool calls. It is not an interface the agent navigates where APIs are absent.

Sourcedocs.veecle.ai/introductionread 2026-09-23