callback

Get past the ATS gate so you talk to a human recruiter.

callback is a Model Context Protocol (MCP) server for job applications. It fetches job descriptions, scores your resume against ATS keyword requirements, tailors resume bullets using your wiki of behavioral stories, and renders a PDF — all orchestrated by your MCP host (Claude, Codex, etc.).

---

Install

Requires uv — install uv

# 1. Install callback from GitHub
uv tool install git+https://github.com/thedandano/callback.git

# 2. Install Playwright Chromium (needed for URL-based JD fetching)
callback install-browsers

# 3. Register the MCP server with Claude and Codex
callback setup-mcp

After setup-mcp, restart your MCP host (Claude Desktop, etc.) to pick up the new server.

setup-mcp is intentionally noninteractive: it only registers the callback server entry for Claude and Codex. It preserves any existing env map on the callback server entry, but it does not ask for API keys or configure tracing. Server logs default to ~/.local/state/callback/server.log. Project-local logs are opt-in for manual debugging with callback serve --project-logs.

For local development installs from this repository, prefer:

make install && callback setup-mcp

make install embeds a local build version in the CLI. Clean installs on origin/main print the package version, while local commits ahead of origin/main print a suffix like 0.3.0-01-abc1234; uncommitted changes add -dirty.

Install as a plugin

This is the recommended path for most users — it is self-contained: you do not need to install the callback package separately. The plugin's MCP server bootstraps itself with uvx directly from GitHub, so installing the plugin is all you need.

Requires uv on your PATH (install uv). First launch is slower: uvx fetches dependencies and the server auto-installs Playwright Chromium (used for JD fetching and PDF rendering) on first start.

Two ways in: the polished plugin flow on Claude Code, or a paste-in config for any other MCP client.

Claude Code

/plugin marketplace add thedandano/callback
/plugin install callback@callback

Or the one-shot CLI wrapper (needs the claude CLI on PATH):

callback setup-plugin --target claude      # add --print-only to preview the commands

For local development, install from a clone instead of GitHub:

callback setup-plugin --target claude --plugin-source /path/to/callback

Any other MCP client (Codex, Cursor, Claude Desktop, …)

Add this to the client's MCP config — uvx downloads, builds, and runs the server straight from the GitHub URL, with no separate install:

{
  "mcpServers": {
    "callback": {
      "command": "uvx",
      "args": [
        "--from",
        "git+https://github.com/thedandano/callback",
        "callback",
        "serve"
      ]
    }
  }
}

Restart the client afterward. Codex users can alternatively run codex marketplace add github:thedandano/callback then codex plugin add callback@callback.

A published PyPI release (so this becomes uvx callback serve) is planned once the GitHub-URL install is proven out.

After install

After installing the plugin, restart your Claude Code or Codex session, or run /reload-plugins to activate the MCP server. The setup-plugin command requires claude and/or codex to be on your PATH for the chosen target.

The MCP server runs via uvx (see .mcp.json) and needs no separate package install. On its first start it auto-installs Playwright Chromium in the background; the --skip-browsers flag above only affects setup-plugin's pre-warm step, not this server-side bootstrap.

Config

MCP hosts launch callback serve as a subprocess, so tracing environment variables must live in the Claude/Codex MCP config env maps.

# Guided LangSmith setup for both Claude and Codex
callback config langsmith

# Noninteractive LangSmith setup
callback config langsmith --api-key lsv2-... --project Callback

# Inspect configured env vars. Secret-like keys are redacted by default.
callback config status
callback config env list
callback config env list --show-secrets

# Set or unset one env var
callback config env set CALLBACK_TRACE_BACKEND langsmith --target all
callback config env unset LANGSMITH_API_KEY --target codex

# Verify active LangSmith configuration and optionally emit a safe test span
callback trace-check --target codex
callback trace-check --target codex --emit-test-trace

LangSmith tracing is opt-in. To enable it, configure:

• CALLBACK_TRACE_BACKEND=langsmith
• LANGSMITH_TRACING=true
• LANGSMITH_ENDPOINT=https://api.smith.langchain.com
• LANGSMITH_API_KEY=<your LangSmith API key>
• LANGSMITH_PROJECT=Callback unless you choose another project name

LangSmith offers hosted accounts, including free-tier signup paths. callback does not require LangSmith to run; without the env vars above, tracing is a no-op and normal logs still work.

callback uses two tracing layers when LangSmith is enabled:

• LangGraph RunnableConfig metadata for graph runs.
• Sanitized LangSmith decorator spans for MCP tool calls and graph nodes.

The MCP server suppresses native LangChain/LangGraph auto-tracing around graph invocation because callback intentionally pauses graphs at host handoff points. Without this, the LangSmith UI can show a native graph run as pending while it is waiting for the next MCP tool call. The sanitized callback.* tool and node spans are the supported observability surface for MCP demos.

The decorator spans are intentionally redacted. They include safe fields such as session_id, tool_name, graph_name, node_name, resume_label, counts, and key names. They do not include resume text, JD body text, wiki page content, host-submitted edits, file paths, or API keys.

After any callback config ... change, restart Claude/Codex so the MCP host reloads the subprocess environment.

Use callback config status before or after setup-mcp to compare the Claude/Codex callback env maps. It is read-only and reports same, different, missing, or unset so you can confirm existing env vars were not overwritten.

Logs

callback logs --follow

callback logs reads the default ~/.local/state/callback/server.log. Use --project-logs to read ./.callback/server.log, or --log-path <path> to override both.

---

Update

callback update

Upgrades to the latest release via uv tool upgrade.

---

Uninstall

# Step 1: remove MCP server entries from Claude and Codex configs
callback uninstall

# Step 2: remove the tool itself
uv tool uninstall callback

Add --purge to step 1 to also delete application data and state directories:

callback uninstall --purge

--purge deletes ~/.local/share/callback/ (application PDFs, JSON archives, and LangGraph SQLite checkpointer databases) and ~/.local/state/callback/ (logs).

---

How to use (MCP tools)

Once the server is running, your MCP host can call these tools:

Tool	Description
load_jd	Fetch a job description from a URL or accept raw text. Returns JD markdown and extraction instructions.
submit_keywords	Submit host-extracted JDData (title, required/preferred keywords, seniority, years). Scores your resume and returns keyword gaps.
submit_tailor	Apply bullet edits to your resume SectionMap and finalize — renders PDF, scores final result, produces report.
get_wiki_pages	Fetch pages from your behavioral-story wiki by path.
onboard_user	Register a new resume, skills file, and accomplishments doc.
compile_profile	Recompile your profile from all stored stories.
create_story	Persist a new behavioral story (SBI format) for a skill.
check_update	Return current version, latest GitHub release tag, and update_available flag.

Agent MCP Playbook

When a user asks to use callback for a job, the MCP host should follow the workflow metadata returned by each tool:

1. Call load_jd with jd_url or jd_raw_text.
2. Extract compact JDData from data.jd_text using data.extraction_protocol.
3. Call submit_keywords with the same session_id and the compact jd_json string.
4. If workflow.next_tool is get_wiki_pages, use data.wiki_index to fetch relevant evidence pages, then call submit_tailor.
5. If workflow.next_tool is submit_tailor, create honest edits from data.sections, data.score_gap, wiki evidence, and data.tailor_instructions.
6. If workflow.next_tool is onboard_user or create_story, collect the missing profile evidence, compile the profile, then restart the job flow with load_jd.
7. After submit_tailor, return data.pdf_path, data.archive_path, data.report, and data.outcome to the user.

The host owns keyword extraction and tailoring judgment. callback owns state, validation, rendering, scoring, and archival.

Scoring dimensions

Dimension	Max pts	Signal
KeywordMatch	45	Required (×0.7) + preferred (×0.3) keywords
ExperienceFit	25	Years met + seniority match
ImpactEvidence	10	Quantified metric bullets
ATSFormat	10	Standard section headers present
Readability	10	Absence of filler phrases

---

Development

# Clone and install with dev dependencies
git clone https://github.com/thedandano/callback.git
cd callback
uv sync

# One-time browser setup
uv run playwright install chromium

# Run tests
uv run pytest

# Type check
uv run pyright

# Run the MCP server locally
uv run python -m callback.server

Scheduling a recurring scan

Point your scheduler at the scan-job-leads skill — no bespoke automation skill is needed.

Claude /schedule example (runs every weekday morning):

/schedule "0 8 * * 1-5" Use the scan-job-leads skill. Read search preferences via get_search_preferences, scan configured sources within lead_recency_days, dedupe against the ledger, and return the standard scan-job-leads summary with recommended leads for apply-to-job.

cron example (replace claude with your Claude Code CLI invocation):

0 8 * * 1-5  claude --print "Use the scan-job-leads skill." >> ~/job-scan.log 2>&1

After a scan, run apply-to-job on selected full-source leads to score, tailor, and stage applications — submission always requires explicit current-turn approval.