Ambient Context is a macOS menu bar app that watches what you work on and turns it into something your coding agent can actually use. It reads the focused window through the Accessibility API every few seconds, builds a cited knowledge base from that record, and writes a daily summary. The whole thing runs locally, stores everything as plain markdown, and ships with a 20-tool MCP server so agents can read and drive it directly.
What It Captures and What It Produces
Each day yields three outputs, visible as tabs in the app's day view. Context is the raw record: what was on screen, when, in which app, and the document path or page URL where the app exposes one. Knowledge is a small wiki built by an agent you connect, with six files covering People, Commitments, Threads, Products, Issues, and Reading, each claim cited back to a block in the record. Notes is a written summary produced from the timeline headings and knowledge base rather than raw bodies, with citations checked against the evidence.
Capture never reads background windows, never runs while the screen is locked, and takes no screenshots, video, or OCR. It reads the focused window's text through macOS Accessibility, deduplicates lines across the day, and strips tracking parameters from URLs. Two full days of real capture, about 12,500 lines, showed that 22% of recorded content was bare counters and social chrome: view counts, vote counts, "8 minutes ago" strings, media player positions ticking every poll. The noise filter drops counter-shaped lines and pipe-separated navigation menus, producing roughly 10% smaller output with no real content lost.
Agent Skills Install in One Click
The v1.1.0 release added five agent skills bundled inside the app binary. The core skill teaches agents what the record is, which tool answers which question, and how to handle restraint and privacy. Four workflow skills cover catching up on missed days, standups, weekly reviews, and tuning capture rules.
Installing writes them to ~/.claude/skills/ and ~/.agents/skills/, which Claude Code, Cursor, Codex, Zed, Copilot, Gemini CLI, Goose, and opencode all read. The app tracks what it wrote and never overwrites or deletes an ASKILL.md file you have edited without explicit confirmation. An Update button appears when a release changes a skill.
Self-Healing Login Items and Startup Diagnostics
The v1.1.2 release fixed a subtle login item bug. Turning launch at login on from a development build under target/ wrote a login item naming that binary. Every login since started the stale dev build beside the installed app because the startup check only asked whether a login item existed, not whether it pointed to the right binary. The check now reads the program the login item names and rewrites it when the installed app finds it pointing elsewhere.
A panic during startup now writes to a panic.log file in the app's data folder with the thread and source location, since startup panics happen inside a system callback that cannot unwind and the standard crash report only records the abort.
A 20-Tool MCP Server With an Audit Trail
The app binary doubles as a stdio MCP server via the mcp subcommand. Read tools open the capture folder directly and work whether or not the app is running. Tools that change anything go through the running app over a local Unix socket at mode 0600, so validation and the ledger happen in one place and every write carries the client's name.
The tool set covers record operations (list days, read day, search, open), knowledge and notes (read KB, read summary, ingest day, summarize day), capture control (status, start, stop), rules (list, add, update, remove), configuration (get/set config, get/set prompt), and a ledger that records every run, who triggered it, how long it took, and any agent output that was rejected.
Processing can run on a schedule, once a day at a time you choose, or on demand from the window or over MCP. The pipeline runs on an agent CLI you already have. Ambient Context holds no credentials and makes no model calls of its own.
Privacy by Design and a Documented Threat Model
There is no Ambient Context server, no account, no telemetry, and no upload of the record. The optional update check reaches GitHub Releases and nothing else. Before a block is written, secure password fields are skipped at the Accessibility source, recognized password managers and private browsing windows are dropped, and known credential, API-key, and card-shaped patterns are scrubbed to [redacted].
Two hardening changes protect readers from a record that lies to the model. A body line that looks like a block heading cannot forge a heading: the writer escapes such lines, and readers accept a heading only in heading position. A summary whose citations or figures are not in the evidence is rejected rather than saved. The full trust boundary, data inventory, and known gaps are documented in the repo.
The project is MIT licensed, requires macOS 14 or later on Apple Silicon, and available on GitHub. It has been in active development since the first working version on August 25, 2026, with signed and notarized builds that preserve the Accessibility grant across updates.