Skip to content

TruthGraph · The context layer for AI coding

Your agent can readevery file,or just the right ones.

20% faster answers from Codex, Claude Code, and Cursor.
30% more usage out of every model subscription. Internal measurements; results vary by task and model.

Start 14-day free trial See the 3-step setup

No card required · 14 days free · No automatic charges

  • Codex
  • Claude Code
  • Cursor
  • Any MCP-capable agent
Built around the code you actually ship.

Source Relationships Evidence Revision

From zero to context

Set it up in 60 seconds.

Three steps. No infra. No configuration. One copy-paste.

  1. 01

    Register

    Create your TruthGraph account. 14-day trial, no card.

    Create account
  2. 02

    Add a PAT key

    Paste a scoped GitHub or GitLab token in Account → Providers. Your agent reuses it without ever seeing it.

    Add a provider token
  3. 03

    Paste the prompt into Codex or Claude

    Copy our ready-made install prompt into your coding agent. It installs the skill, connects your repositories, and starts answering with real code context.

    Get the install prompt

The 60-second tour

See how context condenses.

A question goes in. The graph lights up. Focused source comes back.

01 · Agent asks

What depends on POST /checkout?

03 · Context packet
  • services/checkout.tshandler
  • analytics/track.tssubscriber
  • billing/webhook.tscaller
  • contracts/checkout.jsonschema
@ rev 7f3a1cevidence attached

Looping preview · animated SVG, no video required

The source stays attached

Useful to the model.
Checkable by you.

Every answer carries its source references and the indexed revision. Follow modeled dependencies across repositories. Direct evidence and inferred relationships stay distinct, with confidence shown for inference.

  • Source location + revision on every fact
  • Inferred edges labeled, scored, and separable
  • Cross-repo closure for impact questions
Inspect the graph

Recorded internal tests · September 2026

Seconds to context.

See what we measured
100%

Top-five retrieval hit rate

11 of 11 questions found the expected source in the top five, across 8 synthetic source documents.

Encoder-ranking evaluation
0.25–1.31s

Observed hosted MCP response time

Seven recorded calls covering exact lookup, natural-language discovery, and dependency impact.

Existing operator deployment
~40%

Fewer tokens / task

Our estimate from development usage. Savings vary by task, model, conversation length, and how your agent manages context.

Estimated · internal measurement

In our development workflow, Codex could spend several minutes on repeated grep searches and file reads. TruthGraph gives agents a focused retrieval step, with source references they can follow.

How we measured these results

Internal tests from September 30, 2026. The accuracy fixture used 11 paraphrased questions and 8 synthetic documents: 10 expected sources ranked first and one ranked third. It measured encoder ranking, not full-repository or end-to-end MCP accuracy.

The seven hosted MCP calls took 253–1,306 ms. Cache state was not reset; the document worker was active and source imports were paused. These are separate samples, not a combined accuracy-and-speed guarantee or a service-level promise.

The Codex search comparison describes our development experience, not a controlled head-to-head test. Results vary with the question, indexed coverage, workload, and deployment.

Two ways in. Evidence either way.

Give your model a useful starting point.

A question needs more than a matching file. Bring the surrounding relationships and source references into the answer.

01 / AGENT-LED

Let your agent retrieve.

Install the skill and give your agent a key. It can retrieve relevant code, relationships, and evidence. With your optional repository-management permission, it can also connect repositories, configure scans, and check indexing. Revoke access any time.

AgentTruthGraph skillContext
Get the skill & setup prompt
02 / PREASSEMBLED

Put context in the prompt.

Your runner retrieves the bundle, loads the selected source, and adds it to the model's prompt. The model can use that supplied context without discovering those files or making retrieval calls itself.

Bundle + sourcePromptLocal model

Each bundle has a scope. A task may need additional retrieval.

Keep the context current

Connect once.
Index the branch you ship from.

  1. 01

    Connect a repository.

    Choose GitHub or GitLab and authorize repository access.

  2. 02

    Choose your branch.

    Index the branch you work from, so retrieval has an explicit revision.

  3. 03

    Stay up to date.

    Install a push hook to queue a fresh scan when that branch changes.

Context for the next change

Give your next agent
a clearer starting point.

Start your 14-day free trial

No card required · 14 days free · No automatic charges