Embedder

Embedder acts in GitHub through a GitHub App: it commits, pushes a branch, opens or updates a pull request and posts review findings on changed lines, and from Slack it can do the same and post a completion summary in the thread. Customers add their own tools as MCP servers over stdio, HTTP or SSE, with credentials passed as headers or a browser sign in, set globally or per project, and can hide individual tools or flag the ones that control live hardware.

On the bench it drives J-Link, OpenOCD, probe-rs and TI CCS debuggers, UART, RTT, SWO and CAN transports, Saleae and Digilent analyzers, PicoScope and Siglent scopes, Joulescope and Nordic PPK2 power meters and BLE devices. A remote bridge reaches serial ports, probes and flash tools on another machine over an authenticated connection.

Sourcedocs.embedder.com/integrations/mcp-servers.mdread 2026-10-08

Work can split across several agents. Read only subagents (codebase explorer, document explorer, web searcher and planner) run in parallel, and a general subagent can run in the background, return a task ID, take messages, be interrupted without losing its transcript and report back when done. Hardware access stays off for subagents unless someone turns it on.

Goals keep the main agent working toward one outcome across turns until it completes, hits a limit or needs input, and Plan, Act and Debug modes separate a reviewed plan from execution and investigation. A hardware queue keeps two sessions from driving the same equipment at once, and the daemon runs up to four GitHub and Slack tasks in parallel by default.

Sourcedocs.embedder.com/core-concepts/subagents.mdread 2026-10-08

Each project keeps a structured model of the customer's own board. Uploaded KiCad, Altium, Eagle, PADS, Xpedition or EDIF designs are turned into components, pins, nets, power connections, buses, sheets, power domains and functional blocks, and the agent can be pointed at any of them in chat, such as a component U1 or a net +3V3. A new board revision can replace the design or amend it by file, and Embedder compares components, nets and pins with the previous version.

Around that model, an EMBEDDER.md file records the target, build and test commands, debug setup, bench equipment and project rules, and the hardware catalog supplies board documentation and skills for supported parts. Before changing a driver or a pin assignment, the agent checks its assumptions against the design.

Sourcedocs.embedder.com/core-concepts/schematics.mdread 2026-10-08

Three approval modes set when the agent asks first: Always ask, Manually approve, the default, which lets ordinary reads and approved commands run and asks before anything else, and Skip all approvals. Each request can be allowed once, allowed for the session or rejected. Command prefix rules, a project approvals file that can only request permissions rather than grant them, and a .embedderignore list of protected files narrow what runs without a prompt.

Plan mode produces a reviewed approach before any change, and dispatching a subagent needs consent. Starting a goal switches on automatic approvals, so the person writes any prohibited changes into the goal and can pause or clear it at any time. In GitHub, merging a pull request takes a separate request, and admins can pause automatic reviews and autofix for an installation.

Sourcedocs.embedder.com/core-concepts/approvals.mdread 2026-10-08

Embedder states a SOC 2 Type II attestation covering security, availability and confidentiality, ISO 27001 certification with a formal ISMS and DPA support, and says Enterprise inputs are not used to train models. Uploads of shared conversations redact detected secrets, and the remote hardware bridge pins its TLS key, stores client tokens hashed, locks out repeated failed sign ins, can restrict which programs run to an allowlist and audits every execution.

Inside the customer's organization, the picture is thinner. Users sign in through VS Code and join teams, and the docs refer to organization owner, organization admin and team admin roles for connecting GitHub and Slack, but no single sign on, SCIM or permission settings are described.

Sourceembedder.comread 2026-10-08

Every conversation keeps a saved transcript of the agent's work, searchable in project history and shareable with teammates as a read only snapshot with secrets redacted. GitHub and Slack tasks report the pull request, branch and commit they produced, hardware scripts return JSON tied to the script and firmware revision, and trace sessions are saved with their events and instructions. The Startup plan adds centralized usage and spend analytics, and the remote bridge audits every command it runs.

Embedder does not describe a team wide audit log of who did what across users and agents.

Sourcedocs.embedder.com/core-concepts/sessions.mdread 2026-10-08

Project memory holds confirmed conventions, constraints, hardware issues and decisions, shared across every conversation in the project. Entries last until someone deletes them, and clearing a conversation does not remove them. A project holds up to 50, and five of them can be pinned to load at the start of each conversation while the rest stay searchable.

The agent proposes an entry when asked to remember something, the save goes through the approval settings, and entries can be listed, updated, pinned or forgotten on request, with deletion done in Act mode. Temporary observations and unfinished task notes stay in the conversation instead, and conversations themselves can be resumed or forked later.

Sourcedocs.embedder.com/core-concepts/memory.mdread 2026-10-08

Customers pick the deployment: Embedder's secure cloud, a single tenant deployment on Enterprise, or fully on premises or in their own cloud so source code and schematics stay in infrastructure they control. The CLI connects to a self hosted instance with embedder connect, the GitHub and Slack daemon runs on the customer's own machine with their toolchain, and a remote bridge can run as a Linux or macOS service on a lab host, reached through Embedder's relay rather than an open network port.

Sourceembedder.com/pricingread 2026-10-08

Skills ship bundled: debug-gdb, debug-saleae, debug-digilent and debug-picoscope for captures, debug-joulescope and debug-ppk2 for power, dashboard-authoring and optimize, validate-MISRA and validate-CERT for static analysis, and router skills for USB CDC and Ethernet drivers that hand off to platform skills for STM32, ESP32, TinyUSB, NXP, Renesas, Zephyr and Linux. Each handles its own job.

Four read only subagents come built in, and the hardware catalog adds board documentation and skills for more than 500 platforms. Customers install more skills from a SKILL.md folder, a file or an HTTPS source, and Embedder also picks up skills kept in .agents/skills or .claude/skills folders.

Sourcedocs.embedder.com/core-concepts/skills.mdread 2026-10-08

Work reaches Embedder from outside the editor. Mentioning @embedder in a connected GitHub repository, or mentioning the Embedder app in a Slack channel, hands a task to a daemon on the customer's machine, and replies in the same thread answer its questions or resume a paused task. Per repository reviews run on pull requests on their own, with an optional autofix run.

Inside a session, loops repeat a reviewed task every 5, 15 or 60 minutes for 3, 10 or 24 runs, expiring after three days. Loops can read code, docs and the web, update plans and make scoped edits, but they do not run shell commands, flash hardware or start live captures.

Sourcedocs.embedder.com/headless/github.mdread 2026-10-08

The customer chooses the model. Embedder's hosted models are on by default and can be switched off, GitHub Copilot and ChatGPT Plus or Pro subscriptions connect with /connect, and any local OpenAI compatible server, such as Ollama, LM Studio, vLLM or llama.cpp, can be selected. A project file pins a main and a fast model when the project opens, and the free plan offers a narrower set of models than Startup.

Sourcedocs.embedder.com/integrations/model-providers.mdread 2026-10-08

Outside callers reach Embedder through its CLI and daemon. The daemon authenticates unattended with an API key, takes work from GitHub and Slack, and runs tasks in parallel, and hardware scripts return their result as JSON with embedder script --json. The CLI also manages MCP servers, skills, dashboards and remote bridges from a terminal or an SSH session.

Embedder publishes no REST API, SDK or MCP server of its own. Its MCP support connects other servers to Embedder rather than letting other agents drive it.

Sourcedocs.embedder.com/headless/cli-reference.mdread 2026-10-08

Changes are checked against explicit pass criteria. Hardware scripts are Python files in .embedder/hardware/ that pair setup, a bounded measurement, a pass condition and cleanup, run through Embedder's runner with shared helpers for analyzers and scopes, and return JSON with a success flag, a summary and supporting observations. They run from chat, VS Code or embedder script --json, with per bench profiles. A smoke test passes only when the expected marker appears within its time limit, and a missing result is treated as a failure.

The agent adds regression cases tied to a failure and reruns the same check after a fix. SEGGER J-Trace reports function and line coverage for a workload, and static analysis runs through cppcheck, clang-tidy, Perforce QAC, Parasoft or Coverity, with validate-MISRA and validate-CERT reporting a scan as complete only when the analyzer finished on the whole scope.

Sourcedocs.embedder.com/solutions/automated-testing.mdread 2026-10-08

Embedder operates lab instruments, debug probes and serial ports through their own drivers and tools, and a Siglent scope screenshot comes over the network rather than from a screen the agent drives. A web searcher subagent looks things up online, and there is no browser or desktop session operated by the agent.

Sourcedocs.embedder.com/core-concepts/subagents.mdread 2026-10-08